clarity

* update props

* update documents

* nit

* update chat processing

* k

* k

* nit

* minor nit

* minor nits

* k

* nits

backend/danswer/agent_search/answer_query/graph_builder.py

@@ -1,100 +0,0 @@
-from langgraph.graph import END
-from langgraph.graph import START
-from langgraph.graph import StateGraph
-
-from danswer.agent_search.answer_query.nodes.answer_check import answer_check
-from danswer.agent_search.answer_query.nodes.answer_generation import answer_generation
-from danswer.agent_search.answer_query.nodes.format_answer import format_answer
-from danswer.agent_search.answer_query.states import AnswerQueryInput
-from danswer.agent_search.answer_query.states import AnswerQueryOutput
-from danswer.agent_search.answer_query.states import AnswerQueryState
-from danswer.agent_search.expanded_retrieval.graph_builder import (
-    expanded_retrieval_graph_builder,
-)
-
-
-def answer_query_graph_builder() -> StateGraph:
-    graph = StateGraph(
-        state_schema=AnswerQueryState,
-        input=AnswerQueryInput,
-        output=AnswerQueryOutput,
-    )
-
-    ### Add nodes ###
-
-    expanded_retrieval = expanded_retrieval_graph_builder().compile()
-    graph.add_node(
-        node="expanded_retrieval_for_initial_decomp",
-        action=expanded_retrieval,
-    )
-    graph.add_node(
-        node="answer_check",
-        action=answer_check,
-    )
-    graph.add_node(
-        node="answer_generation",
-        action=answer_generation,
-    )
-    graph.add_node(
-        node="format_answer",
-        action=format_answer,
-    )
-
-    ### Add edges ###
-
-    graph.add_edge(
-        start_key=START,
-        end_key="expanded_retrieval_for_initial_decomp",
-    )
-    graph.add_edge(
-        start_key="expanded_retrieval_for_initial_decomp",
-        end_key="answer_generation",
-    )
-    graph.add_edge(
-        start_key="answer_generation",
-        end_key="answer_check",
-    )
-    graph.add_edge(
-        start_key="answer_check",
-        end_key="format_answer",
-    )
-    graph.add_edge(
-        start_key="format_answer",
-        end_key=END,
-    )
-
-    return graph
-
-
-if __name__ == "__main__":
-    from danswer.db.engine import get_session_context_manager
-    from danswer.llm.factory import get_default_llms
-    from danswer.context.search.models import SearchRequest
-
-    graph = answer_query_graph_builder()
-    compiled_graph = graph.compile()
-    primary_llm, fast_llm = get_default_llms()
-    search_request = SearchRequest(
-        query="Who made Excel and what other products did they make?",
-    )
-    with get_session_context_manager() as db_session:
-        inputs = AnswerQueryInput(
-            search_request=search_request,
-            primary_llm=primary_llm,
-            fast_llm=fast_llm,
-            db_session=db_session,
-            query_to_answer="Who made Excel?",
-        )
-        output = compiled_graph.invoke(
-            input=inputs,
-            # debug=True,
-            # subgraphs=True,
-        )
-        print(output)
-        # for namespace, chunk in compiled_graph.stream(
-        #     input=inputs,
-        #     # debug=True,
-        #     subgraphs=True,
-        # ):
-        #     print(namespace)
-        #     print(chunk)

backend/danswer/agent_search/answer_query/nodes/answer_check.py

@@ -1,30 +0,0 @@
-from langchain_core.messages import HumanMessage
-from langchain_core.messages import merge_message_runs
-
-from danswer.agent_search.answer_query.states import AnswerQueryState
-from danswer.agent_search.answer_query.states import QACheckOutput
-from danswer.agent_search.shared_graph_utils.prompts import BASE_CHECK_PROMPT
-
-
-def answer_check(state: AnswerQueryState) -> QACheckOutput:
-    msg = [
-        HumanMessage(
-            content=BASE_CHECK_PROMPT.format(
-                question=state["search_request"].query,
-                base_answer=state["answer"],
-            )
-        )
-    ]
-
-    fast_llm = state["fast_llm"]
-    response = list(
-        fast_llm.stream(
-            prompt=msg,
-        )
-    )
-
-    response_str = merge_message_runs(response, chunk_separator="")[0].content
-
-    return QACheckOutput(
-        answer_quality=response_str,
-    )

backend/danswer/agent_search/answer_query/nodes/answer_generation.py

@@ -1,32 +0,0 @@
-from langchain_core.messages import HumanMessage
-from langchain_core.messages import merge_message_runs
-
-from danswer.agent_search.answer_query.states import AnswerQueryState
-from danswer.agent_search.answer_query.states import QAGenerationOutput
-from danswer.agent_search.shared_graph_utils.prompts import BASE_RAG_PROMPT
-from danswer.agent_search.shared_graph_utils.utils import format_docs
-
-
-def answer_generation(state: AnswerQueryState) -> QAGenerationOutput:
-    query = state["query_to_answer"]
-    docs = state["reranked_documents"]
-
-    print(f"Number of verified retrieval docs: {len(docs)}")
-
-    msg = [
-        HumanMessage(
-            content=BASE_RAG_PROMPT.format(question=query, context=format_docs(docs))
-        )
-    ]
-
-    fast_llm = state["fast_llm"]
-    response = list(
-        fast_llm.stream(
-            prompt=msg,
-        )
-    )
-
-    answer_str = merge_message_runs(response, chunk_separator="")[0].content
-    return QAGenerationOutput(
-        answer=answer_str,
-    )

backend/danswer/agent_search/answer_query/nodes/format_answer.py

@@ -1,16 +0,0 @@
-from danswer.agent_search.answer_query.states import AnswerQueryOutput
-from danswer.agent_search.answer_query.states import AnswerQueryState
-from danswer.agent_search.answer_query.states import SearchAnswerResults
-
-
-def format_answer(state: AnswerQueryState) -> AnswerQueryOutput:
-    return AnswerQueryOutput(
-        decomp_answer_results=[
-            SearchAnswerResults(
-                query=state["query_to_answer"],
-                quality=state["answer_quality"],
-                answer=state["answer"],
-                documents=state["reranked_documents"],
-            )
-        ],
-    )

backend/danswer/agent_search/answer_query/states.py

@@ -1,48 +0,0 @@
-from typing import Annotated
-from typing import TypedDict
-
-from pydantic import BaseModel
-
-from danswer.agent_search.core_state import PrimaryState
-from danswer.agent_search.shared_graph_utils.operators import dedup_inference_sections
-from danswer.context.search.models import InferenceSection
-
-
-class SearchAnswerResults(BaseModel):
-    query: str
-    answer: str
-    quality: str
-    documents: Annotated[list[InferenceSection], dedup_inference_sections]
-
-
-class QACheckOutput(TypedDict, total=False):
-    answer_quality: str
-
-
-class QAGenerationOutput(TypedDict, total=False):
-    answer: str
-
-
-class ExpandedRetrievalOutput(TypedDict):
-    reranked_documents: Annotated[list[InferenceSection], dedup_inference_sections]
-
-
-class AnswerQueryState(
-    PrimaryState,
-    QACheckOutput,
-    QAGenerationOutput,
-    ExpandedRetrievalOutput,
-    total=True,
-):
-    query_to_answer: str
-    retrieved_documents: Annotated[list[InferenceSection], dedup_inference_sections]
-    verified_documents: Annotated[list[InferenceSection], dedup_inference_sections]
-    reranked_documents: Annotated[list[InferenceSection], dedup_inference_sections]
-
-
-class AnswerQueryInput(PrimaryState, total=True):
-    query_to_answer: str
-
-
-class AnswerQueryOutput(TypedDict):
-    decomp_answer_results: list[SearchAnswerResults]

backend/danswer/agent_search/core_state.py

@@ -1,15 +0,0 @@
-from typing import TypedDict
-
-from sqlalchemy.orm import Session
-
-from danswer.context.search.models import SearchRequest
-from danswer.llm.interfaces import LLM
-
-
-class PrimaryState(TypedDict, total=False):
-    search_request: SearchRequest
-    primary_llm: LLM
-    fast_llm: LLM
-    # a single session for the entire agent search
-    # is fine if we are only reading
-    db_session: Session

backend/danswer/agent_search/deep_answer/nodes/answer_generation.py

@@ -1,114 +0,0 @@
-from typing import Any
-
-from langchain_core.messages import HumanMessage
-
-from danswer.agent_search.main.states import MainState
-from danswer.agent_search.shared_graph_utils.prompts import COMBINED_CONTEXT
-from danswer.agent_search.shared_graph_utils.prompts import MODIFIED_RAG_PROMPT
-from danswer.agent_search.shared_graph_utils.utils import format_docs
-from danswer.agent_search.shared_graph_utils.utils import normalize_whitespace
-
-
-# aggregate sub questions and answers
-def deep_answer_generation(state: MainState) -> dict[str, Any]:
-    """
-    Generate answer
-
-    Args:
-        state (messages): The current state
-
-    Returns:
-         dict: The updated state with re-phrased question
-    """
-    print("---DEEP GENERATE---")
-
-    question = state["original_question"]
-    docs = state["deduped_retrieval_docs"]
-
-    deep_answer_context = state["core_answer_dynamic_context"]
-
-    print(f"Number of verified retrieval docs - deep: {len(docs)}")
-
-    combined_context = normalize_whitespace(
-        COMBINED_CONTEXT.format(
-            deep_answer_context=deep_answer_context, formated_docs=format_docs(docs)
-        )
-    )
-
-    msg = [
-        HumanMessage(
-            content=MODIFIED_RAG_PROMPT.format(
-                question=question, combined_context=combined_context
-            )
-        )
-    ]
-
-    # Grader
-    model = state["fast_llm"]
-    response = model.invoke(msg)
-
-    return {
-        "deep_answer": response.content,
-    }
-
-
-def final_stuff(state: MainState) -> dict[str, Any]:
-    """
-    Invokes the agent model to generate a response based on the current state. Given
-    the question, it will decide to retrieve using the retriever tool, or simply end.
-
-    Args:
-        state (messages): The current state
-
-    Returns:
-        dict: The updated state with the agent response appended to messages
-    """
-    print("---FINAL---")
-
-    messages = state["log_messages"]
-    time_ordered_messages = [x.pretty_repr() for x in messages]
-    time_ordered_messages.sort()
-
-    print("Message Log:")
-    print("\n".join(time_ordered_messages))
-
-    initial_sub_qas = state["initial_sub_qas"]
-    initial_sub_qa_list = []
-    for initial_sub_qa in initial_sub_qas:
-        if initial_sub_qa["sub_answer_check"] == "yes":
-            initial_sub_qa_list.append(
-                f'  Question:\n  {initial_sub_qa["sub_question"]}\n  --\n  Answer:\n  {initial_sub_qa["sub_answer"]}\n  -----'
-            )
-
-    initial_sub_qa_context = "\n".join(initial_sub_qa_list)
-
-    base_answer = state["base_answer"]
-
-    print(f"Final Base Answer:\n{base_answer}")
-    print("--------------------------------")
-    print(f"Initial Answered Sub Questions:\n{initial_sub_qa_context}")
-    print("--------------------------------")
-
-    if not state.get("deep_answer"):
-        print("No Deep Answer was required")
-        return {}
-
-    deep_answer = state["deep_answer"]
-    sub_qas = state["sub_qas"]
-    sub_qa_list = []
-    for sub_qa in sub_qas:
-        if sub_qa["sub_answer_check"] == "yes":
-            sub_qa_list.append(
-                f'  Question:\n  {sub_qa["sub_question"]}\n  --\n  Answer:\n  {sub_qa["sub_answer"]}\n  -----'
-            )
-
-    sub_qa_context = "\n".join(sub_qa_list)
-
-    print(f"Final Base Answer:\n{base_answer}")
-    print("--------------------------------")
-    print(f"Final Deep Answer:\n{deep_answer}")
-    print("--------------------------------")
-    print("Sub Questions and Answers:")
-    print(sub_qa_context)
-
-    return {}

backend/danswer/agent_search/deep_answer/nodes/deep_decomp.py

@@ -1,78 +0,0 @@
-import json
-import re
-from datetime import datetime
-from typing import Any
-
-from langchain_core.messages import HumanMessage
-
-from danswer.agent_search.main.states import MainState
-from danswer.agent_search.shared_graph_utils.prompts import DEEP_DECOMPOSE_PROMPT
-from danswer.agent_search.shared_graph_utils.utils import format_entity_term_extraction
-from danswer.agent_search.shared_graph_utils.utils import generate_log_message
-
-
-def decompose(state: MainState) -> dict[str, Any]:
-    """ """
-
-    node_start_time = datetime.now()
-
-    question = state["original_question"]
-    base_answer = state["base_answer"]
-
-    # get the entity term extraction dict and properly format it
-    entity_term_extraction_dict = state["retrieved_entities_relationships"][
-        "retrieved_entities_relationships"
-    ]
-
-    entity_term_extraction_str = format_entity_term_extraction(
-        entity_term_extraction_dict
-    )
-
-    initial_question_answers = state["initial_sub_qas"]
-
-    addressed_question_list = [
-        x["sub_question"]
-        for x in initial_question_answers
-        if x["sub_answer_check"] == "yes"
-    ]
-    failed_question_list = [
-        x["sub_question"]
-        for x in initial_question_answers
-        if x["sub_answer_check"] == "no"
-    ]
-
-    msg = [
-        HumanMessage(
-            content=DEEP_DECOMPOSE_PROMPT.format(
-                question=question,
-                entity_term_extraction_str=entity_term_extraction_str,
-                base_answer=base_answer,
-                answered_sub_questions="\n - ".join(addressed_question_list),
-                failed_sub_questions="\n - ".join(failed_question_list),
-            ),
-        )
-    ]
-
-    # Grader
-    model = state["fast_llm"]
-    response = model.invoke(msg)
-
-    cleaned_response = re.sub(r"```json\n|\n```", "", response.pretty_repr())
-    parsed_response = json.loads(cleaned_response)
-
-    sub_questions_dict = {}
-    for sub_question_nr, sub_question_dict in enumerate(
-        parsed_response["sub_questions"]
-    ):
-        sub_question_dict["answered"] = False
-        sub_question_dict["verified"] = False
-        sub_questions_dict[sub_question_nr] = sub_question_dict
-
-    return {
-        "decomposed_sub_questions_dict": sub_questions_dict,
-        "log_messages": generate_log_message(
-            message="deep - decompose",
-            node_start_time=node_start_time,
-            graph_start_time=state["graph_start_time"],
-        ),
-    }

backend/danswer/agent_search/deep_answer/nodes/entity_term_extraction.py

@@ -1,40 +0,0 @@
-import json
-import re
-from typing import Any
-
-from langchain_core.messages import HumanMessage
-from langchain_core.messages import merge_message_runs
-
-from danswer.agent_search.main.states import MainState
-from danswer.agent_search.shared_graph_utils.prompts import ENTITY_TERM_PROMPT
-from danswer.agent_search.shared_graph_utils.utils import format_docs
-
-
-def entity_term_extraction(state: MainState) -> dict[str, Any]:
-    """Extract entities and terms from the question and context"""
-
-    question = state["original_question"]
-    docs = state["deduped_retrieval_docs"]
-
-    doc_context = format_docs(docs)
-
-    msg = [
-        HumanMessage(
-            content=ENTITY_TERM_PROMPT.format(question=question, context=doc_context),
-        )
-    ]
-    fast_llm = state["fast_llm"]
-    # Grader
-    llm_response_list = list(
-        fast_llm.stream(
-            prompt=msg,
-        )
-    )
-    llm_response = merge_message_runs(llm_response_list, chunk_separator="")[0].content
-
-    cleaned_response = re.sub(r"```json\n|\n```", "", llm_response)
-    parsed_response = json.loads(cleaned_response)
-
-    return {
-        "retrieved_entities_relationships": parsed_response,
-    }

backend/danswer/agent_search/deep_answer/nodes/sub_qa_level_aggregator.py

@@ -1,30 +0,0 @@
-from typing import Any
-
-from danswer.agent_search.main.states import MainState
-
-
-# aggregate sub questions and answers
-def sub_qa_level_aggregator(state: MainState) -> dict[str, Any]:
-    sub_qas = state["sub_qas"]
-
-    dynamic_context_list = [
-        "Below you will find useful information to answer the original question:"
-    ]
-    checked_sub_qas = []
-
-    for core_answer_sub_qa in sub_qas:
-        question = core_answer_sub_qa["sub_question"]
-        answer = core_answer_sub_qa["sub_answer"]
-        verified = core_answer_sub_qa["sub_answer_check"]
-
-        if verified == "yes":
-            dynamic_context_list.append(
-                f"Question:\n{question}\n\nAnswer:\n{answer}\n\n---\n\n"
-            )
-            checked_sub_qas.append({"sub_question": question, "sub_answer": answer})
-    dynamic_context = "\n".join(dynamic_context_list)
-
-    return {
-        "core_answer_dynamic_context": dynamic_context,
-        "checked_sub_qas": checked_sub_qas,
-    }

backend/danswer/agent_search/deep_answer/nodes/sub_qa_manager.py

@@ -1,19 +0,0 @@
-from typing import Any
-
-from danswer.agent_search.main.states import MainState
-
-
-def sub_qa_manager(state: MainState) -> dict[str, Any]:
-    """ """
-
-    sub_questions_dict = state["decomposed_sub_questions_dict"]
-
-    sub_questions = {}
-
-    for sub_question_nr, sub_question_dict in sub_questions_dict.items():
-        sub_questions[sub_question_nr] = sub_question_dict["sub_question"]
-
-    return {
-        "sub_questions": sub_questions,
-        "num_new_question_iterations": 0,
-    }

backend/danswer/agent_search/expanded_retrieval/edges.py

@@ -1,83 +0,0 @@
-from collections.abc import Hashable
-
-from langchain_core.messages import HumanMessage
-from langchain_core.messages import merge_message_runs
-from langgraph.types import Send
-
-from danswer.agent_search.expanded_retrieval.nodes.doc_retrieval import RetrieveInput
-from danswer.agent_search.expanded_retrieval.states import DocRetrievalOutput
-from danswer.agent_search.expanded_retrieval.states import DocVerificationInput
-from danswer.agent_search.expanded_retrieval.states import ExpandedRetrievalInput
-from danswer.agent_search.shared_graph_utils.prompts import (
-    REWRITE_PROMPT_MULTI_ORIGINAL,
-)
-from danswer.llm.interfaces import LLM
-
-
-def parallel_retrieval_edge(state: ExpandedRetrievalInput) -> list[Send | Hashable]:
-    # print(f"parallel_retrieval_edge state: {state.keys()}")
-    print("parallel_retrieval_edge state")
-
-    # This should be better...
-    question = state.get("query_to_answer") or state["search_request"].query
-    llm: LLM = state["fast_llm"]
-
-    """
-    msg = [
-        HumanMessage(
-            content=REWRITE_PROMPT_MULTI.format(question=question),
-        )
-    ]
-    """
-    msg = [
-        HumanMessage(
-            content=REWRITE_PROMPT_MULTI_ORIGINAL.format(question=question),
-        )
-    ]
-
-    llm_response_list = list(
-        llm.stream(
-            prompt=msg,
-        )
-    )
-    llm_response = merge_message_runs(llm_response_list, chunk_separator="")[0].content
-
-    # print(f"llm_response: {llm_response}")
-
-    rewritten_queries = [
-        rewritten_query.strip() for rewritten_query in llm_response.split("--")
-    ]
-
-    # Add the original sub-question as one of the 'rewritten' queries
-    rewritten_queries = [question] + rewritten_queries
-
-    print(f"rewritten_queries: {rewritten_queries}")
-
-    return [
-        Send(
-            "doc_retrieval",
-            RetrieveInput(query_to_retrieve=query, **state),
-        )
-        for query in rewritten_queries
-    ]
-
-
-def parallel_verification_edge(state: DocRetrievalOutput) -> list[Send | Hashable]:
-    # print(f"parallel_retrieval_edge state: {state.keys()}")
-    print("parallel_retrieval_edge state")
-
-    retrieved_docs = state["retrieved_documents"]
-
-    return [
-        Send(
-            "doc_verification",
-            DocVerificationInput(doc_to_verify=doc, **state),
-        )
-        for doc in retrieved_docs
-    ]
-
-
-# this is not correct - remove
-# def conditionally_rerank_edge(state: ExpandedRetrievalState) -> bool:
-#    print(f"conditionally_rerank_edge state: {state.keys()}")
-#    return bool(state["search_request"].rerank_settings)

backend/danswer/agent_search/expanded_retrieval/graph_builder.py

@@ -1,129 +0,0 @@
-from langgraph.graph import END
-from langgraph.graph import START
-from langgraph.graph import StateGraph
-
-from danswer.agent_search.expanded_retrieval.edges import parallel_retrieval_edge
-from danswer.agent_search.expanded_retrieval.edges import parallel_verification_edge
-from danswer.agent_search.expanded_retrieval.nodes.doc_reranking import doc_reranking
-from danswer.agent_search.expanded_retrieval.nodes.doc_retrieval import doc_retrieval
-from danswer.agent_search.expanded_retrieval.nodes.doc_verification import (
-    doc_verification,
-)
-from danswer.agent_search.expanded_retrieval.nodes.dummy_node import dummy_node
-from danswer.agent_search.expanded_retrieval.nodes.verification_kickoff import (
-    verification_kickoff,
-)
-from danswer.agent_search.expanded_retrieval.states import ExpandedRetrievalInput
-from danswer.agent_search.expanded_retrieval.states import ExpandedRetrievalOutput
-from danswer.agent_search.expanded_retrieval.states import ExpandedRetrievalState
-
-# from danswer.agent_search.expanded_retrieval.edges import conditionally_rerank_edge
-
-
-def expanded_retrieval_graph_builder() -> StateGraph:
-    graph = StateGraph(
-        state_schema=ExpandedRetrievalState,
-        input=ExpandedRetrievalInput,
-        output=ExpandedRetrievalOutput,
-    )
-
-    ### Add nodes ###
-
-    graph.add_node(
-        node="doc_retrieval",
-        action=doc_retrieval,
-    )
-    graph.add_node(
-        node="verification_kickoff",
-        action=verification_kickoff,
-    )
-    graph.add_node(
-        node="doc_verification",
-        action=doc_verification,
-    )
-    graph.add_node(
-        node="doc_reranking",
-        action=doc_reranking,
-    )
-
-    graph.add_node(
-        node="post_retrieval_dummy_node",
-        action=dummy_node,
-    )
-
-    graph.add_node(
-        node="dummy_node",
-        action=dummy_node,
-    )
-
-    ### Add edges ###
-
-    graph.add_conditional_edges(
-        source=START,
-        path=parallel_retrieval_edge,
-        path_map=["doc_retrieval"],
-    )
-
-    graph.add_edge(
-        start_key="doc_retrieval",
-        end_key="verification_kickoff",
-    )
-
-    graph.add_conditional_edges(
-        source="verification_kickoff",
-        path=parallel_verification_edge,
-        path_map=["doc_verification"],
-    )
-
-    # graph.add_edge(
-    #    start_key="doc_verification",
-    #    end_key="post_retrieval_dummy_node",
-    # )
-
-    graph.add_edge(
-        start_key="doc_verification",
-        end_key="doc_reranking",
-    )
-
-    graph.add_edge(
-        start_key="doc_reranking",
-        end_key="dummy_node",
-    )
-
-    # graph.add_conditional_edges(
-    #    source="doc_verification",
-    #    path=conditionally_rerank_edge,
-    #    path_map={
-    #        True: "doc_reranking",
-    #        False: END,
-    #    },
-    # )
-    graph.add_edge(
-        start_key="dummy_node",
-        end_key=END,
-    )
-
-    return graph
-
-
-if __name__ == "__main__":
-    from danswer.db.engine import get_session_context_manager
-    from danswer.llm.factory import get_default_llms
-    from danswer.context.search.models import SearchRequest
-
-    graph = expanded_retrieval_graph_builder()
-    compiled_graph = graph.compile()
-    primary_llm, fast_llm = get_default_llms()
-    search_request = SearchRequest(
-        query="Who made Excel and what other products did they make?",
-    )
-    with get_session_context_manager() as db_session:
-        inputs = ExpandedRetrievalInput(
-            search_request=search_request,
-            primary_llm=primary_llm,
-            fast_llm=fast_llm,
-            db_session=db_session,
-            query_to_answer="Who made Excel?",
-        )
-        for thing in compiled_graph.stream(inputs, debug=True):
-            print(thing)

backend/danswer/agent_search/expanded_retrieval/nodes/doc_reranking.py

@@ -1,13 +0,0 @@
-import datetime
-
-from danswer.agent_search.expanded_retrieval.states import DocRerankingOutput
-from danswer.agent_search.expanded_retrieval.states import ExpandedRetrievalState
-
-
-def doc_reranking(state: ExpandedRetrievalState) -> DocRerankingOutput:
-    print(f"doc_reranking state: {datetime.datetime.now()}")
-
-    verified_documents = state["verified_documents"]
-    reranked_documents = verified_documents
-
-    return DocRerankingOutput(reranked_documents=reranked_documents)

backend/danswer/agent_search/expanded_retrieval/nodes/doc_retrieval.py

@@ -1,75 +0,0 @@
-import datetime
-
-from danswer.agent_search.expanded_retrieval.states import DocRetrievalOutput
-from danswer.agent_search.expanded_retrieval.states import ExpandedRetrievalState
-from danswer.context.search.models import InferenceSection
-from danswer.context.search.models import SearchRequest
-from danswer.context.search.pipeline import SearchPipeline
-
-
-class RetrieveInput(ExpandedRetrievalState):
-    query_to_retrieve: str
-
-
-def doc_retrieval(state: RetrieveInput) -> DocRetrievalOutput:
-    # def doc_retrieval(state: RetrieveInput) -> Command[Literal["doc_verification"]]:
-    """
-    Retrieve documents
-
-    Args:
-        state (dict): The current graph state
-
-    Returns:
-        state (dict): New key added to state, documents, that contains retrieved documents
-    """
-    # print(f"doc_retrieval state: {state.keys()}")
-
-    if "query_to_answer" in state.keys():
-        query_question = state["query_to_answer"]
-    else:
-        query_question = state["search_request"].query
-
-    query_to_retrieve = state["query_to_retrieve"]
-
-    print(f"\ndoc_retrieval state: {datetime.datetime.now()}")
-    print(f"  -- search_request: {query_question[:100]}")
-    # print(f"       -- query_to_retrieve: {query_to_retrieve[:100]}")
-
-    documents: list[InferenceSection] = []
-    llm = state["primary_llm"]
-    fast_llm = state["fast_llm"]
-    # db_session = state["db_session"]
-
-    documents = SearchPipeline(
-        search_request=SearchRequest(
-            query=query_to_retrieve,
-        ),
-        user=None,
-        llm=llm,
-        fast_llm=fast_llm,
-        db_session=state["db_session"],
-    ).reranked_sections
-
-    top_1_score = documents[0].center_chunk.score
-    top_5_score = sum([doc.center_chunk.score for doc in documents[:5]]) / 5
-    top_10_score = sum([doc.center_chunk.score for doc in documents[:10]]) / 10
-
-    fit_score = 1 / 3 * (top_1_score + top_5_score + top_10_score)
-
-    # temp - limit the number of documents to 5
-    documents = documents[:5]
-
-    """
-    chunk_ids = {
-        "query": query_to_retrieve,
-        "chunk_ids": [doc.center_chunk.chunk_id for doc in documents],
-    }
-    """
-
-    print(f"sub_query: {query_to_retrieve[:50]}")
-    print(f"retrieved documents: {len(documents)}")
-    print(f"fit score: {fit_score}")
-    print()
-    return DocRetrievalOutput(
-        retrieved_documents=documents,
-    )

backend/danswer/agent_search/expanded_retrieval/nodes/doc_verification.py

@@ -1,63 +0,0 @@
-import datetime
-
-from langchain_core.messages import HumanMessage
-from langchain_core.messages import merge_message_runs
-
-from danswer.agent_search.expanded_retrieval.states import DocRetrievalOutput
-from danswer.agent_search.expanded_retrieval.states import DocVerificationOutput
-from danswer.agent_search.shared_graph_utils.models import BinaryDecision
-from danswer.agent_search.shared_graph_utils.prompts import VERIFIER_PROMPT
-
-
-def doc_verification(state: DocRetrievalOutput) -> DocVerificationOutput:
-    """
-    Check whether the document is relevant for the original user question
-
-    Args:
-        state (VerifierState): The current state
-
-    Returns:
-        dict: ict: The updated state with the final decision
-    """
-
-    # print(f"--- doc_verification state ---")
-
-    if "query_to_answer" in state.keys():
-        query_to_answer = state["query_to_answer"]
-    else:
-        query_to_answer = state["search_request"].query
-
-    doc_to_verify = state["doc_to_verify"]
-    document_content = doc_to_verify.combined_content
-
-    msg = [
-        HumanMessage(
-            content=VERIFIER_PROMPT.format(
-                question=query_to_answer, document_content=document_content
-            )
-        )
-    ]
-
-    fast_llm = state["fast_llm"]
-    response = list(
-        fast_llm.stream(
-            prompt=msg,
-        )
-    )
-
-    response_string = merge_message_runs(response, chunk_separator="")[0].content
-    # Convert string response to proper dictionary format
-    decision_dict = {"decision": response_string.lower()}
-    formatted_response = BinaryDecision.model_validate(decision_dict)
-
-    verified_documents = []
-    if formatted_response.decision == "yes":
-        verified_documents.append(doc_to_verify)
-
-        print(
-            f"Verdict & Completion: {formatted_response.decision} -- {datetime.datetime.now()}"
-        )
-
-    return DocVerificationOutput(
-        verified_documents=verified_documents,
-    )

backend/danswer/agent_search/expanded_retrieval/nodes/dummy_node.py

@@ -1,9 +0,0 @@
-def dummy_node(state):
-    """
-    This node is a dummy node that does not change the state but allows to inspect the state.
-    """
-    print(f"doc_reranking state: {state.keys()}")
-
-    state["verified_documents"]
-
-    return {}

backend/danswer/agent_search/expanded_retrieval/nodes/verification_kickoff.py

@@ -1,28 +0,0 @@
-import datetime
-from typing import Literal
-
-from langgraph.types import Command
-from langgraph.types import Send
-
-from danswer.agent_search.expanded_retrieval.states import (
-    DocVerificationInput,
-)
-from danswer.agent_search.expanded_retrieval.states import ExpandedRetrievalState
-
-
-def verification_kickoff(
-    state: ExpandedRetrievalState,
-) -> Command[Literal["doc_verification"]]:
-    print(f"verification_kickoff state: {datetime.datetime.now()}")
-
-    documents = state["retrieved_documents"]
-    return Command(
-        update={},
-        goto=[
-            Send(
-                node="doc_verification",
-                arg=DocVerificationInput(doc_to_verify=doc, **state),
-            )
-            for doc in documents
-        ],
-    )

backend/danswer/agent_search/expanded_retrieval/states.py

@@ -1,42 +0,0 @@
-from typing import Annotated
-from typing import TypedDict
-
-from danswer.agent_search.core_state import PrimaryState
-from danswer.agent_search.shared_graph_utils.operators import dedup_inference_sections
-from danswer.context.search.models import InferenceSection
-
-
-class DocRetrievalOutput(TypedDict, total=False):
-    retrieved_documents: Annotated[list[InferenceSection], dedup_inference_sections]
-    query_to_answer: str
-
-
-class DocVerificationInput(TypedDict, total=True):
-    query_to_answer: str
-    doc_to_verify: InferenceSection
-
-
-class DocVerificationOutput(TypedDict, total=False):
-    verified_documents: Annotated[list[InferenceSection], dedup_inference_sections]
-
-
-class DocRerankingOutput(TypedDict, total=False):
-    reranked_documents: Annotated[list[InferenceSection], dedup_inference_sections]
-
-
-class ExpandedRetrievalState(
-    PrimaryState,
-    DocRetrievalOutput,
-    DocVerificationOutput,
-    DocRerankingOutput,
-    total=True,
-):
-    query_to_answer: str
-
-
-class ExpandedRetrievalInput(PrimaryState, total=True):
-    query_to_answer: str
-
-
-class ExpandedRetrievalOutput(TypedDict):
-    reranked_documents: Annotated[list[InferenceSection], dedup_inference_sections]

backend/danswer/agent_search/main/edges.py

@@ -1,61 +0,0 @@
-from collections.abc import Hashable
-
-from langgraph.types import Send
-
-from danswer.agent_search.answer_query.states import AnswerQueryInput
-from danswer.agent_search.main.states import MainState
-
-
-def parallelize_decompozed_answer_queries(state: MainState) -> list[Send | Hashable]:
-    return [
-        Send(
-            "answer_query",
-            AnswerQueryInput(
-                **state,
-                query_to_answer=query,
-            ),
-        )
-        for query in state["initial_decomp_queries"]
-    ]
-
-
-# def continue_to_answer_sub_questions(state: QAState) -> Union[Hashable, list[Hashable]]:
-#     # Routes re-written queries to the (parallel) retrieval steps
-#     # Notice the 'Send()' API that takes care of the parallelization
-#     return [
-#         Send(
-#             "sub_answers_graph",
-#             ResearchQAState(
-#                 sub_question=sub_question["sub_question_str"],
-#                 sub_question_nr=sub_question["sub_question_nr"],
-#                 graph_start_time=state["graph_start_time"],
-#                 primary_llm=state["primary_llm"],
-#                 fast_llm=state["fast_llm"],
-#             ),
-#         )
-#         for sub_question in state["sub_questions"]
-#     ]
-
-
-# def continue_to_deep_answer(state: QAState) -> Union[Hashable, list[Hashable]]:
-#     print("---GO TO DEEP ANSWER OR END---")
-
-#     base_answer = state["base_answer"]
-
-#     question = state["original_question"]
-
-#     BASE_CHECK_MESSAGE = [
-#         HumanMessage(
-#             content=BASE_CHECK_PROMPT.format(question=question, base_answer=base_answer)
-#         )
-#     ]
-
-#     model = state["fast_llm"]
-#     response = model.invoke(BASE_CHECK_MESSAGE)
-
-#     print(f"CAN WE CONTINUE W/O GENERATING A DEEP ANSWER? - {response.pretty_repr()}")
-
-#     if response.pretty_repr() == "no":
-#         return "decompose"
-#     else:
-#         return "end"

backend/danswer/agent_search/main/graph_builder.py

@@ -1,125 +0,0 @@
-import datetime
-
-from langgraph.graph import END
-from langgraph.graph import START
-from langgraph.graph import StateGraph
-
-from danswer.agent_search.answer_query.graph_builder import answer_query_graph_builder
-from danswer.agent_search.expanded_retrieval.graph_builder import (
-    expanded_retrieval_graph_builder,
-)
-from danswer.agent_search.main.edges import parallelize_decompozed_answer_queries
-from danswer.agent_search.main.nodes.base_decomp import main_decomp_base
-from danswer.agent_search.main.nodes.dummy_node import dummy_node
-from danswer.agent_search.main.nodes.generate_initial_answer import (
-    generate_initial_answer,
-)
-from danswer.agent_search.main.states import MainInput
-from danswer.agent_search.main.states import MainState
-
-
-def main_graph_builder() -> StateGraph:
-    graph = StateGraph(
-        state_schema=MainState,
-        input=MainInput,
-    )
-
-    ### Add nodes ###
-
-    graph.add_node(
-        node="dummy_node_start",
-        action=dummy_node,
-    )
-
-    graph.add_node(
-        node="dummy_node_right",
-        action=dummy_node,
-    )
-
-    graph.add_node(
-        node="base_decomp",
-        action=main_decomp_base,
-    )
-    answer_query_subgraph = answer_query_graph_builder().compile()
-    graph.add_node(
-        node="answer_query",
-        action=answer_query_subgraph,
-    )
-    expanded_retrieval_subgraph = expanded_retrieval_graph_builder().compile()
-    graph.add_node(
-        node="expanded_retrieval",
-        action=expanded_retrieval_subgraph,
-    )
-    graph.add_node(
-        node="generate_initial_answer",
-        action=generate_initial_answer,
-    )
-
-    ### Add edges ###
-
-    graph.add_edge(
-        start_key=START,
-        end_key="dummy_node_start",
-    )
-
-    graph.add_edge(
-        start_key="dummy_node_start",
-        end_key="dummy_node_right",
-    )
-    graph.add_edge(
-        start_key="dummy_node_right",
-        end_key="expanded_retrieval",
-    )
-    # graph.add_edge(
-    #    start_key="expanded_retrieval",
-    #    end_key="generate_initial_answer",
-    # )
-
-    graph.add_edge(
-        start_key="dummy_node_start",
-        end_key="base_decomp",
-    )
-    graph.add_conditional_edges(
-        source="base_decomp",
-        path=parallelize_decompozed_answer_queries,
-        path_map=["answer_query"],
-    )
-    graph.add_edge(
-        start_key=["answer_query", "expanded_retrieval"],
-        end_key="generate_initial_answer",
-    )
-    graph.add_edge(
-        start_key="generate_initial_answer",
-        end_key=END,
-    )
-
-    return graph
-
-
-if __name__ == "__main__":
-    from danswer.db.engine import get_session_context_manager
-    from danswer.llm.factory import get_default_llms
-    from danswer.context.search.models import SearchRequest
-
-    graph = main_graph_builder()
-    compiled_graph = graph.compile()
-    primary_llm, fast_llm = get_default_llms()
-    search_request = SearchRequest(
-        query="Who made Excel and what other products did they make?",
-    )
-    with get_session_context_manager() as db_session:
-        inputs = MainInput(
-            search_request=search_request,
-            primary_llm=primary_llm,
-            fast_llm=fast_llm,
-            db_session=db_session,
-        )
-
-        print(f"START: {datetime.datetime.now()}")
-
-        output = compiled_graph.invoke(
-            input=inputs,
-            # debug=True,
-            # subgraphs=True,
-        )
-        print(output)

backend/danswer/agent_search/main/nodes/base_decomp.py

@@ -1,35 +0,0 @@
-import datetime
-
-from langchain_core.messages import HumanMessage
-
-from danswer.agent_search.main.states import BaseDecompOutput
-from danswer.agent_search.main.states import MainState
-from danswer.agent_search.shared_graph_utils.prompts import INITIAL_DECOMPOSITION_PROMPT
-from danswer.agent_search.shared_graph_utils.utils import clean_and_parse_list_string
-
-
-def main_decomp_base(state: MainState) -> BaseDecompOutput:
-    print(f"main_decomp_base state: {datetime.datetime.now()}")
-    question = state["search_request"].query
-
-    msg = [
-        HumanMessage(
-            content=INITIAL_DECOMPOSITION_PROMPT.format(question=question),
-        )
-    ]
-
-    # Get the rewritten queries in a defined format
-    model = state["fast_llm"]
-    response = model.invoke(msg)
-
-    content = response.pretty_repr()
-    list_of_subquestions = clean_and_parse_list_string(content)
-
-    decomp_list: list[str] = [
-        sub_question["sub_question"].strip() for sub_question in list_of_subquestions
-    ]
-
-    print(f"Decomp Questions: {decomp_list}")
-    return BaseDecompOutput(
-        initial_decomp_queries=decomp_list,
-    )

backend/danswer/agent_search/main/nodes/dummy_node.py

@@ -1,10 +0,0 @@
-import datetime
-
-
-def dummy_node(state):
-    """
-    This node is a dummy node that does not change the state but allows to inspect the state.
-    """
-    print(f"DUMMY NODE: {datetime.datetime.now()}")
-
-    return {}

backend/danswer/agent_search/main/nodes/generate_initial_answer.py

@@ -1,51 +0,0 @@
-from langchain_core.messages import HumanMessage
-
-from danswer.agent_search.main.states import InitialAnswerOutput
-from danswer.agent_search.main.states import MainState
-from danswer.agent_search.shared_graph_utils.prompts import INITIAL_RAG_PROMPT
-from danswer.agent_search.shared_graph_utils.utils import format_docs
-
-
-def generate_initial_answer(state: MainState) -> InitialAnswerOutput:
-    print("---GENERATE INITIAL---")
-
-    question = state["search_request"].query
-    docs = state["documents"]
-
-    decomp_answer_results = state["decomp_answer_results"]
-
-    good_qa_list: list[str] = []
-
-    _SUB_QUESTION_ANSWER_TEMPLATE = """
-    Sub-Question:\n  - {sub_question}\n  --\nAnswer:\n  - {sub_answer}\n\n
-    """
-    for decomp_answer_result in decomp_answer_results:
-        if (
-            decomp_answer_result.quality == "yes"
-            and len(decomp_answer_result.answer) > 0
-            and decomp_answer_result.answer != "I don't know"
-        ):
-            good_qa_list.append(
-                _SUB_QUESTION_ANSWER_TEMPLATE.format(
-                    sub_question=decomp_answer_result.query,
-                    sub_answer=decomp_answer_result.answer,
-                )
-            )
-
-    sub_question_answer_str = "\n\n------\n\n".join(good_qa_list)
-
-    msg = [
-        HumanMessage(
-            content=INITIAL_RAG_PROMPT.format(
-                question=question,
-                context=format_docs(docs),
-                answered_sub_questions=sub_question_answer_str,
-            )
-        )
-    ]
-
-    # Grader
-    model = state["fast_llm"]
-    response = model.invoke(msg)
-
-    return InitialAnswerOutput(initial_answer=response.pretty_repr())

backend/danswer/agent_search/main/states.py

@@ -1,37 +0,0 @@
-from operator import add
-from typing import Annotated
-from typing import TypedDict
-
-from danswer.agent_search.answer_query.states import SearchAnswerResults
-from danswer.agent_search.core_state import PrimaryState
-from danswer.agent_search.shared_graph_utils.operators import dedup_inference_sections
-from danswer.context.search.models import InferenceSection
-
-
-class BaseDecompOutput(TypedDict, total=False):
-    initial_decomp_queries: list[str]
-
-
-class InitialAnswerOutput(TypedDict, total=False):
-    initial_answer: str
-
-
-class MainState(
-    PrimaryState,
-    BaseDecompOutput,
-    InitialAnswerOutput,
-    total=True,
-):
-    documents: Annotated[list[InferenceSection], dedup_inference_sections]
-    decomp_answer_results: Annotated[list[SearchAnswerResults], add]
-
-
-class MainInput(PrimaryState, total=True):
-    pass
-
-
-class MainOutput(TypedDict):
-    """
-    This is not used because defining the output only matters for filtering the output of
-      a .invoke() call but we are streaming so we just yield the entire state.
-    """

backend/danswer/agent_search/run_graph.py

@@ -1,27 +0,0 @@
-from danswer.agent_search.primary_graph.graph_builder import build_core_graph
-from danswer.llm.answering.answer import AnswerStream
-from danswer.llm.interfaces import LLM
-from danswer.tools.tool import Tool
-
-
-def run_graph(
-    query: str,
-    llm: LLM,
-    tools: list[Tool],
-) -> AnswerStream:
-    graph = build_core_graph()
-
-    inputs = {
-        "original_query": query,
-        "messages": [],
-        "tools": tools,
-        "llm": llm,
-    }
-    compiled_graph = graph.compile()
-    output = compiled_graph.invoke(input=inputs)
-    yield from output
-
-
-if __name__ == "__main__":
-    pass
-    # run_graph("What is the capital of France?", llm, [])

backend/danswer/agent_search/shared_graph_utils/models.py

@@ -1,12 +0,0 @@
-from typing import Literal
-
-from pydantic import BaseModel
-
-
-# Pydantic models for structured outputs
-class RewrittenQueries(BaseModel):
-    rewritten_queries: list[str]
-
-
-class BinaryDecision(BaseModel):
-    decision: Literal["yes", "no"]

backend/danswer/agent_search/shared_graph_utils/operators.py

@@ -1,9 +0,0 @@
-from danswer.context.search.models import InferenceSection
-from danswer.llm.answering.prune_and_merge import _merge_sections
-
-
-def dedup_inference_sections(
-    list1: list[InferenceSection], list2: list[InferenceSection]
-) -> list[InferenceSection]:
-    deduped = _merge_sections(list1 + list2)
-    return deduped

backend/danswer/agent_search/shared_graph_utils/prompts.py

@@ -1,427 +0,0 @@
-REWRITE_PROMPT_MULTI_ORIGINAL = """ \n
-    Please convert an initial user question into a 2-3 more appropriate short and pointed search queries for retrievel from a
-    document store. Particularly, try to think about resolving ambiguities and make the search queries more specific,
-    enabling the system to search more broadly.
-    Also, try to make the search queries not redundant, i.e. not too similar! \n\n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-    Formulate the queries separated by '--' (Do not say 'Query 1: ...', just write the querytext): """
-
-REWRITE_PROMPT_MULTI = """ \n
-    Please create a list of 2-3 sample documents that could answer an original question. Each document
-    should be about as long as the original question. \n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-    Formulate the sample documents separated by '--' (Do not say 'Document 1: ...', just write the text): """
-
-BASE_RAG_PROMPT = """ \n
-    You are an assistant for question-answering tasks. Use the context provided below - and only the
-    provided context - to answer the question. If you don't know the answer or if the provided context is
-    empty, just say "I don't know". Do not use your internal knowledge!
-
-    Again, only use the provided context and do not use your internal knowledge! If you cannot answer the
-    question based on the context, say "I don't know". It is a matter of life and death that you do NOT
-    use your internal knowledge, just the provided information!
-
-    Use three sentences maximum and keep the answer concise.
-    answer concise.\nQuestion:\n {question} \nContext:\n {context} \n\n
-    \n\n
-    Answer:"""
-
-BASE_CHECK_PROMPT = """ \n
-    Please check whether 1) the suggested answer seems to fully address the original question AND 2)the
-    original question requests a simple, factual answer, and there are no ambiguities, judgements,
-    aggregations, or any other complications that may require extra context. (I.e., if the question is
-    somewhat addressed, but the answer would benefit from more context, then answer with 'no'.)
-
-    Please only answer with 'yes' or 'no' \n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-    Here is the proposed answer:
-    \n ------- \n
-    {base_answer}
-    \n ------- \n
-    Please answer with yes or no:"""
-
-VERIFIER_PROMPT = """ \n
-    Please check whether the document seems to be relevant for the answer of the question. Please
-    only answer with 'yes' or 'no' \n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-    Here is the document text:
-    \n ------- \n
-    {document_content}
-    \n ------- \n
-    Please answer with yes or no:"""
-
-INITIAL_DECOMPOSITION_PROMPT_BASIC = """ \n
-    Please decompose an initial user question into not more than 4 appropriate sub-questions that help to
-    answer the original question. The purpose for this decomposition is to isolate individulal entities
-    (i.e., 'compare sales of company A and company B' -> 'what are sales for company A' + 'what are sales
-    for company B'), split ambiguous terms (i.e., 'what is our success with company A' -> 'what are our
-    sales with company A' + 'what is our market share with company A' + 'is company A a reference customer
-    for us'), etc. Each sub-question should be realistically be answerable by a good RAG system. \n
-
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-
-    Please formulate your answer as a list of subquestions:
-
-    Answer:
-    """
-
-REWRITE_PROMPT_SINGLE = """ \n
-    Please convert an initial user question into a more appropriate search query for retrievel from a
-    document store. \n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-
-    Formulate the query: """
-
-MODIFIED_RAG_PROMPT = """You are an assistant for question-answering tasks. Use the context provided below
-    - and only this context - to answer the question. If you don't know the answer, just say "I don't know".
-    Use three sentences maximum and keep the answer concise.
-    Pay also particular attention to the sub-questions and their answers, at least it may enrich the answer.
-    Again, only use the provided context and do not use your internal knowledge! If you cannot answer the
-    question based on the context, say "I don't know". It is a matter of life and death that you do NOT
-    use your internal knowledge, just the provided information!
-
-    \nQuestion: {question}
-    \nContext: {combined_context} \n
-
-    Answer:"""
-
-ORIG_DEEP_DECOMPOSE_PROMPT = """ \n
-    An initial user question needs to be answered. An initial answer has been provided but it wasn't quite
-    good enough. Also, some sub-questions had been answered and this information has been used to provide
-    the initial answer. Some other subquestions may have been suggested based on little knowledge, but they
-    were not directly answerable. Also, some entities, relationships and terms are givenm to you so that
-    you have an idea of how the avaiolable data looks like.
-
-    Your role is to generate 3-5 new sub-questions that would help to answer the initial question,
-    considering:
-
-    1) The initial question
-    2) The initial answer that was found to be unsatisfactory
-    3) The sub-questions that were answered
-    4) The sub-questions that were suggested but not answered
-    5) The entities, relationships and terms that were extracted from the context
-
-    The individual questions should be answerable by a good RAG system.
-    So a good idea would be to use the sub-questions to resolve ambiguities and/or to separate the
-    question for different entities that may be involved in the original question, but in a way that does
-    not duplicate questions that were already tried.
-
-    Additional Guidelines:
-    - The sub-questions should be specific to the question and provide richer context for the question,
-    resolve ambiguities, or address shortcoming of the initial answer
-    - Each sub-question - when answered - should be relevant for the answer to the original question
-    - The sub-questions should be free from comparisions, ambiguities,judgements, aggregations, or any
-    other complications that may require extra context.
-    - The sub-questions MUST have the full context of the original question so that it can be executed by
-    a RAG system independently without the original question available
-      (Example:
-        - initial question: "What is the capital of France?"
-        - bad sub-question: "What is the name of the river there?"
-        - good sub-question: "What is the name of the river that flows through Paris?"
-    - For each sub-question, please provide a short explanation for why it is a good sub-question. So
-    generate a list of dictionaries with the following format:
-      [{{"sub_question": <sub-question>, "explanation": <explanation>, "search_term": <rewrite the
-      sub-question using as a search phrase for the document store>}}, ...]
-
-    \n\n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-
-    Here is the initial sub-optimal answer:
-    \n ------- \n
-    {base_answer}
-    \n ------- \n
-
-    Here are the sub-questions that were answered:
-    \n ------- \n
-    {answered_sub_questions}
-    \n ------- \n
-
-    Here are the sub-questions that were suggested but not answered:
-    \n ------- \n
-    {failed_sub_questions}
-    \n ------- \n
-
-    And here are the entities, relationships and terms extracted from the context:
-    \n ------- \n
-    {entity_term_extraction_str}
-    \n ------- \n
-
-   Please generate the list of good, fully contextualized sub-questions that would help to address the
-   main question. Again, please find questions that are NOT overlapping too much with the already answered
-   sub-questions or those that already were suggested and failed.
-   In other words - what can we try in addition to what has been tried so far?
-
-   Please think through it step by step and then generate the list of json dictionaries with the following
-   format:
-
-   {{"sub_questions": [{{"sub_question": <sub-question>,
-        "explanation": <explanation>,
-        "search_term": <rewrite the sub-question using as a search phrase for the document store>}},
-        ...]}} """
-
-DEEP_DECOMPOSE_PROMPT = """ \n
-    An initial user question needs to be answered. An initial answer has been provided but it wasn't quite
-    good enough. Also, some sub-questions had been answered and this information has been used to provide
-    the initial answer. Some other subquestions may have been suggested based on little knowledge, but they
-    were not directly answerable. Also, some entities, relationships and terms are givenm to you so that
-    you have an idea of how the avaiolable data looks like.
-
-    Your role is to generate 4-6 new sub-questions that would help to answer the initial question,
-    considering:
-
-    1) The initial question
-    2) The initial answer that was found to be unsatisfactory
-    3) The sub-questions that were answered
-    4) The sub-questions that were suggested but not answered
-    5) The entities, relationships and terms that were extracted from the context
-
-    The individual questions should be answerable by a good RAG system.
-    So a good idea would be to use the sub-questions to resolve ambiguities and/or to separate the
-    question for different entities that may be involved in the original question, but in a way that does
-    not duplicate questions that were already tried.
-
-    Additional Guidelines:
-    - The sub-questions should be specific to the question and provide richer context for the question,
-    resolve ambiguities, or address shortcoming of the initial answer
-    - Each sub-question - when answered - should be relevant for the answer to the original question
-    - The sub-questions should be free from comparisions, ambiguities,judgements, aggregations, or any
-    other complications that may require extra context.
-    - The sub-questions MUST have the full context of the original question so that it can be executed by
-    a RAG system independently without the original question available
-      (Example:
-        - initial question: "What is the capital of France?"
-        - bad sub-question: "What is the name of the river there?"
-        - good sub-question: "What is the name of the river that flows through Paris?"
-    - For each sub-question, please also provide a search term that can be used to retrieve relevant
-    documents from a document store.
-    \n\n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-
-    Here is the initial sub-optimal answer:
-    \n ------- \n
-    {base_answer}
-    \n ------- \n
-
-    Here are the sub-questions that were answered:
-    \n ------- \n
-    {answered_sub_questions}
-    \n ------- \n
-
-    Here are the sub-questions that were suggested but not answered:
-    \n ------- \n
-    {failed_sub_questions}
-    \n ------- \n
-
-    And here are the entities, relationships and terms extracted from the context:
-    \n ------- \n
-    {entity_term_extraction_str}
-    \n ------- \n
-
-   Please generate the list of good, fully contextualized sub-questions that would help to address the
-   main question. Again, please find questions that are NOT overlapping too much with the already answered
-   sub-questions or those that already were suggested and failed.
-   In other words - what can we try in addition to what has been tried so far?
-
-   Generate the list of json dictionaries with the following format:
-
-   {{"sub_questions": [{{"sub_question": <sub-question>,
-        "search_term": <rewrite the sub-question using as a search phrase for the document store>}},
-        ...]}} """
-
-DECOMPOSE_PROMPT = """ \n
-    For an initial user question, please generate at 5-10 individual sub-questions whose answers would help
-    \n to answer the initial question. The individual questions should be answerable by a good RAG system.
-    So a good idea would be to \n use the sub-questions to resolve ambiguities and/or to separate the
-    question for different entities that may be involved in the original question.
-
-    In order to arrive at meaningful sub-questions, please also consider the context retrieved from the
-    document store, expressed as entities, relationships and terms. You can also think about the types
-    mentioned in brackets
-
-    Guidelines:
-    - The sub-questions should be specific to the question and provide richer context for the question,
-    and or resolve ambiguities
-    - Each sub-question - when answered - should be relevant for the answer to the original question
-    - The sub-questions should be free from comparisions, ambiguities,judgements, aggregations, or any
-    other complications that may require extra context.
-    - The sub-questions MUST have the full context of the original question so that it can be executed by
-    a RAG system independently without the original question available
-      (Example:
-        - initial question: "What is the capital of France?"
-        - bad sub-question: "What is the name of the river there?"
-        - good sub-question: "What is the name of the river that flows through Paris?"
-    - For each sub-question, please provide a short explanation for why it is a good sub-question. So
-    generate a list of dictionaries with the following format:
-      [{{"sub_question": <sub-question>, "explanation": <explanation>, "search_term": <rewrite the
-      sub-question using as a search phrase for the document store>}}, ...]
-
-    \n\n
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-
-    And here are the entities, relationships and terms extracted from the context:
-    \n ------- \n
-    {entity_term_extraction_str}
-    \n ------- \n
-
-   Please generate the list of good, fully contextualized sub-questions that would help to address the
-   main question. Don't be too specific unless the original question is specific.
-   Please think through it step by step and then generate the list of json dictionaries with the following
-   format:
-   {{"sub_questions": [{{"sub_question": <sub-question>,
-        "explanation": <explanation>,
-        "search_term": <rewrite the sub-question using as a search phrase for the document store>}},
-        ...]}} """
-
-#### Consolidations
-COMBINED_CONTEXT = """-------
-    Below you will find useful information to answer the original question. First, you see a number of
-    sub-questions with their answers. This information should be considered to be more focussed and
-    somewhat more specific to the original question as it tries to contextualized facts.
-    After that will see the documents that were considered to be relevant to answer the original question.
-
-    Here are the sub-questions and their answers:
-    \n\n {deep_answer_context} \n\n
-    \n\n Here are the documents that were considered to be relevant to answer the original question:
-    \n\n {formated_docs} \n\n
-    ----------------
-    """
-
-SUB_QUESTION_EXPLANATION_RANKER_PROMPT = """-------
-    Below you will find a question that we ultimately want to answer (the original question) and a list of
-    motivations in arbitrary order for generated sub-questions that are supposed to help us answering the
-    original question. The motivations are formatted as <motivation number>:  <motivation explanation>.
-    (Again, the numbering is arbitrary and does not necessarily mean that 1 is the most relevant
-    motivation and 2 is less relevant.)
-
-    Please rank the motivations in order of relevance for answering the original question. Also, try to
-    ensure that the top questions do not duplicate too much, i.e. that they are not too similar.
-    Ultimately, create a list with the motivation numbers where the number of the most relevant
-    motivations comes first.
-
-    Here is the original question:
-    \n\n {original_question} \n\n
-    \n\n Here is the list of sub-question motivations:
-    \n\n {sub_question_explanations} \n\n
-    ----------------
-
-    Please think step by step and then generate the ranked list of motivations.
-
-    Please format your answer as a json object in the following format:
-    {{"reasonning": <explain your reasoning for the ranking>,
-      "ranked_motivations": <ranked list of motivation numbers>}}
-    """
-
-
-INITIAL_DECOMPOSITION_PROMPT = """ \n
-    Please decompose an initial user question into 2 or 3 appropriate sub-questions that help to
-    answer the original question. The purpose for this decomposition is to isolate individulal entities
-    (i.e., 'compare sales of company A and company B' -> 'what are sales for company A' + 'what are sales
-    for company B'), split ambiguous terms (i.e., 'what is our success with company A' -> 'what are our
-    sales with company A' + 'what is our market share with company A' + 'is company A a reference customer
-    for us'), etc. Each sub-question should be realistically be answerable by a good RAG system. \n
-
-    For each sub-question, please also create one search term that can be used to retrieve relevant
-    documents from a document store.
-
-    Here is the initial question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-
-    Please formulate your answer as a list of json objects with the following format:
-
-   [{{"sub_question": <sub-question>, "search_term": <search term>}}, ...]
-
-    Answer:
-    """
-
-INITIAL_RAG_PROMPT = """ \n
-    You are an assistant for question-answering tasks. Use the information provided below - and only the
-    provided information - to answer the provided question.
-
-    The information provided below consists of:
-     1) a number of answered sub-questions - these are very important(!) and definitely should be
-     considered to answer the question.
-     2) a number of documents that were also deemed relevant for the question.
-
-    If you don't know the answer or if the provided information is empty or insufficient, just say
-    "I don't know". Do not use your internal knowledge!
-
-    Again, only use the provided informationand do not use your internal knowledge! It is a matter of life
-    and death that you do NOT use your internal knowledge, just the provided information!
-
-    Try to keep your answer concise.
-
-    And here is the question and the provided information:
-    \n
-    \nQuestion:\n {question}
-
-    \nAnswered Sub-questions:\n {answered_sub_questions}
-
-    \nContext:\n {context} \n\n
-    \n\n
-
-    Answer:"""
-
-ENTITY_TERM_PROMPT = """ \n
-    Based on the original question and the context retieved from a dataset, please generate a list of
-    entities (e.g. companies, organizations, industries, products, locations, etc.), terms and concepts
-    (e.g. sales, revenue, etc.) that are relevant for the question, plus their relations to each other.
-
-    \n\n
-    Here is the original question:
-    \n ------- \n
-    {question}
-    \n ------- \n
-   And here is the context retrieved:
-    \n ------- \n
-    {context}
-    \n ------- \n
-
-    Please format your answer as a json object in the following format:
-
-    {{"retrieved_entities_relationships": {{
-        "entities": [{{
-            "entity_name": <assign a name for the entity>,
-            "entity_type": <specify a short type name for the entity, such as 'company', 'location',...>
-        }}],
-        "relationships": [{{
-            "name": <assign a name for the relationship>,
-            "type": <specify a short type name for the relationship, such as 'sales_to', 'is_location_of',...>,
-            "entities": [<related entity name 1>, <related entity name 2>]
-        }}],
-        "terms": [{{
-            "term_name": <assign a name for the term>,
-            "term_type": <specify a short type name for the term, such as 'revenue', 'market_share',...>,
-            "similar_to": <list terms that are similar to this term>
-        }}]
-    }}
-    }}
-   """

backend/danswer/agent_search/shared_graph_utils/utils.py

@@ -1,101 +0,0 @@
-import ast
-import json
-import re
-from collections.abc import Sequence
-from datetime import datetime
-from datetime import timedelta
-from typing import Any
-
-from danswer.context.search.models import InferenceSection
-
-
-def normalize_whitespace(text: str) -> str:
-    """Normalize whitespace in text to single spaces and strip leading/trailing whitespace."""
-    import re
-
-    return re.sub(r"\s+", " ", text.strip())
-
-
-# Post-processing
-def format_docs(docs: Sequence[InferenceSection]) -> str:
-    return "\n\n".join(doc.combined_content for doc in docs)
-
-
-def clean_and_parse_list_string(json_string: str) -> list[dict]:
-    # Remove any prefixes/labels before the actual JSON content
-    json_string = re.sub(r"^.*?(?=\[)", "", json_string, flags=re.DOTALL)
-
-    # Remove markdown code block markers and any newline prefixes
-    cleaned_string = re.sub(r"```json\n|\n```", "", json_string)
-    cleaned_string = cleaned_string.replace("\\n", " ").replace("\n", " ")
-    cleaned_string = " ".join(cleaned_string.split())
-
-    # Try parsing with json.loads first, fall back to ast.literal_eval
-    try:
-        return json.loads(cleaned_string)
-    except json.JSONDecodeError:
-        try:
-            return ast.literal_eval(cleaned_string)
-        except (ValueError, SyntaxError) as e:
-            raise ValueError(f"Failed to parse JSON string: {cleaned_string}") from e
-
-
-def clean_and_parse_json_string(json_string: str) -> dict[str, Any]:
-    # Remove markdown code block markers and any newline prefixes
-    cleaned_string = re.sub(r"```json\n|\n```", "", json_string)
-    cleaned_string = cleaned_string.replace("\\n", " ").replace("\n", " ")
-    cleaned_string = " ".join(cleaned_string.split())
-    # Parse the cleaned string into a Python dictionary
-    return json.loads(cleaned_string)
-
-
-def format_entity_term_extraction(entity_term_extraction_dict: dict[str, Any]) -> str:
-    entities = entity_term_extraction_dict["entities"]
-    terms = entity_term_extraction_dict["terms"]
-    relationships = entity_term_extraction_dict["relationships"]
-
-    entity_strs = ["\nEntities:\n"]
-    for entity in entities:
-        entity_str = f"{entity['entity_name']} ({entity['entity_type']})"
-        entity_strs.append(entity_str)
-
-    entity_str = "\n - ".join(entity_strs)
-
-    relationship_strs = ["\n\nRelationships:\n"]
-    for relationship in relationships:
-        relationship_str = f"{relationship['name']} ({relationship['type']}): {relationship['entities']}"
-        relationship_strs.append(relationship_str)
-
-    relationship_str = "\n - ".join(relationship_strs)
-
-    term_strs = ["\n\nTerms:\n"]
-    for term in terms:
-        term_str = f"{term['term_name']} ({term['term_type']}): similar to {term['similar_to']}"
-        term_strs.append(term_str)
-
-    term_str = "\n - ".join(term_strs)
-
-    return "\n".join(entity_strs + relationship_strs + term_strs)
-
-
-def _format_time_delta(time: timedelta) -> str:
-    seconds_from_start = f"{((time).seconds):03d}"
-    microseconds_from_start = f"{((time).microseconds):06d}"
-    return f"{seconds_from_start}.{microseconds_from_start}"
-
-
-def generate_log_message(
-    message: str,
-    node_start_time: datetime,
-    graph_start_time: datetime | None = None,
-) -> str:
-    current_time = datetime.now()
-
-    if graph_start_time is not None:
-        graph_time_str = _format_time_delta(current_time - graph_start_time)
-    else:
-        graph_time_str = "N/A"
-
-    node_time_str = _format_time_delta(current_time - node_start_time)
-
-    return f"{graph_time_str} ({node_time_str} s): {message}"

backend/danswer/auth/users.py

@@ -58,7 +58,6 @@ from danswer.auth.schemas import UserRole
 from danswer.auth.schemas import UserUpdate
 from danswer.configs.app_configs import AUTH_TYPE
 from danswer.configs.app_configs import DISABLE_AUTH
-from danswer.configs.app_configs import DISABLE_VERIFICATION
 from danswer.configs.app_configs import EMAIL_FROM
 from danswer.configs.app_configs import REQUIRE_EMAIL_VERIFICATION
 from danswer.configs.app_configs import SESSION_EXPIRE_TIME_SECONDS
@@ -132,11 +131,12 @@ def get_display_email(email: str | None, space_less: bool = False) -> str:
 
 
 def user_needs_to_be_verified() -> bool:
-    # all other auth types besides basic should require users to be
-    # verified
-    return not DISABLE_VERIFICATION and (
-        AUTH_TYPE != AuthType.BASIC or REQUIRE_EMAIL_VERIFICATION
-    )
+    if AUTH_TYPE == AuthType.BASIC:
+        return REQUIRE_EMAIL_VERIFICATION
+
+    # For other auth types, if the user is authenticated it's assumed that
+    # the user is already verified via the external IDP
+    return False
 
 
 def verify_email_is_invited(email: str) -> None:

backend/danswer/chat/answer.py

@@ -6,27 +6,27 @@ from langchain.schema.messages import BaseMessage
 from langchain_core.messages import AIMessageChunk
 from langchain_core.messages import ToolCall
 
+from danswer.chat.llm_response_handler import LLMResponseHandlerManager
 from danswer.chat.models import AnswerQuestionPossibleReturn
+from danswer.chat.models import AnswerStyleConfig
 from danswer.chat.models import CitationInfo
 from danswer.chat.models import DanswerAnswerPiece
-from danswer.file_store.utils import InMemoryChatFile
-from danswer.llm.answering.llm_response_handler import LLMCall
-from danswer.llm.answering.llm_response_handler import LLMResponseHandlerManager
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import PreviousMessage
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
-from danswer.llm.answering.prompts.build import default_build_system_message
-from danswer.llm.answering.prompts.build import default_build_user_message
-from danswer.llm.answering.stream_processing.answer_response_handler import (
+from danswer.chat.models import PromptConfig
+from danswer.chat.prompt_builder.build import AnswerPromptBuilder
+from danswer.chat.prompt_builder.build import default_build_system_message
+from danswer.chat.prompt_builder.build import default_build_user_message
+from danswer.chat.prompt_builder.build import LLMCall
+from danswer.chat.stream_processing.answer_response_handler import (
     CitationResponseHandler,
 )
-from danswer.llm.answering.stream_processing.answer_response_handler import (
+from danswer.chat.stream_processing.answer_response_handler import (
     DummyAnswerResponseHandler,
 )
-from danswer.llm.answering.stream_processing.utils import map_document_id_order
-from danswer.llm.answering.tool.tool_response_handler import ToolResponseHandler
+from danswer.chat.stream_processing.utils import map_document_id_order
+from danswer.chat.tool_handling.tool_response_handler import ToolResponseHandler
+from danswer.file_store.utils import InMemoryChatFile
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.natural_language_processing.utils import get_tokenizer
 from danswer.tools.force import ForceUseTool
 from danswer.tools.models import ToolResponse

backend/danswer/chat/chat_utils.py

@@ -26,7 +26,7 @@ from danswer.db.models import Prompt
 from danswer.db.models import Tool
 from danswer.db.models import User
 from danswer.db.persona import get_prompts_by_ids
-from danswer.llm.answering.models import PreviousMessage
+from danswer.llm.models import PreviousMessage
 from danswer.natural_language_processing.utils import BaseTokenizer
 from danswer.server.query_and_chat.models import CreateChatMessageRequest
 from danswer.tools.tool_implementations.custom.custom_tool import (

backend/danswer/chat/llm_response_handler.py

@@ -1,58 +1,22 @@
 from collections.abc import Callable
 from collections.abc import Generator
 from collections.abc import Iterator
-from typing import TYPE_CHECKING
 
 from langchain_core.messages import BaseMessage
-from pydantic.v1 import BaseModel as BaseModel__v1
 
-from danswer.chat.models import CitationInfo
-from danswer.chat.models import DanswerAnswerPiece
+from danswer.chat.models import ResponsePart
 from danswer.chat.models import StreamStopInfo
 from danswer.chat.models import StreamStopReason
-from danswer.file_store.models import InMemoryChatFile
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
-from danswer.tools.force import ForceUseTool
-from danswer.tools.models import ToolCallFinalResult
-from danswer.tools.models import ToolCallKickoff
-from danswer.tools.models import ToolResponse
-from danswer.tools.tool import Tool
-
-
-if TYPE_CHECKING:
-    from danswer.llm.answering.stream_processing.answer_response_handler import (
-        AnswerResponseHandler,
-    )
-    from danswer.llm.answering.tool.tool_response_handler import ToolResponseHandler
-
-
-ResponsePart = (
-    DanswerAnswerPiece
-    | CitationInfo
-    | ToolCallKickoff
-    | ToolResponse
-    | ToolCallFinalResult
-    | StreamStopInfo
-)
-
-
-class LLMCall(BaseModel__v1):
-    prompt_builder: AnswerPromptBuilder
-    tools: list[Tool]
-    force_use_tool: ForceUseTool
-    files: list[InMemoryChatFile]
-    tool_call_info: list[ToolCallKickoff | ToolResponse | ToolCallFinalResult]
-    using_tool_calling_llm: bool
-
-    class Config:
-        arbitrary_types_allowed = True
+from danswer.chat.prompt_builder.build import LLMCall
+from danswer.chat.stream_processing.answer_response_handler import AnswerResponseHandler
+from danswer.chat.tool_handling.tool_response_handler import ToolResponseHandler
 
 
 class LLMResponseHandlerManager:
     def __init__(
         self,
-        tool_handler: "ToolResponseHandler",
-        answer_handler: "AnswerResponseHandler",
+        tool_handler: ToolResponseHandler,
+        answer_handler: AnswerResponseHandler,
         is_cancelled: Callable[[], bool],
     ):
         self.tool_handler = tool_handler

backend/danswer/chat/models.py

@@ -1,10 +1,14 @@
+from collections.abc import Callable
 from collections.abc import Iterator
 from datetime import datetime
 from enum import Enum
 from typing import Any
+from typing import TYPE_CHECKING
 
 from pydantic import BaseModel
+from pydantic import ConfigDict
 from pydantic import Field
+from pydantic import model_validator
 
 from danswer.configs.constants import DocumentSource
 from danswer.configs.constants import MessageType
@@ -12,8 +16,15 @@ from danswer.context.search.enums import QueryFlow
 from danswer.context.search.enums import RecencyBiasSetting
 from danswer.context.search.enums import SearchType
 from danswer.context.search.models import RetrievalDocs
+from danswer.llm.override_models import PromptOverride
+from danswer.tools.models import ToolCallFinalResult
+from danswer.tools.models import ToolCallKickoff
+from danswer.tools.models import ToolResponse
 from danswer.tools.tool_implementations.custom.base_tool_types import ToolResultType
 
+if TYPE_CHECKING:
+    from danswer.db.models import Prompt
+
 
 class LlmDoc(BaseModel):
     """This contains the minimal set information for the LLM portion including citations"""
@@ -210,3 +221,109 @@ AnswerQuestionStreamReturn = Iterator[AnswerQuestionPossibleReturn]
 class LLMMetricsContainer(BaseModel):
     prompt_tokens: int
     response_tokens: int
+
+
+StreamProcessor = Callable[[Iterator[str]], AnswerQuestionStreamReturn]
+
+
+class DocumentPruningConfig(BaseModel):
+    max_chunks: int | None = None
+    max_window_percentage: float | None = None
+    max_tokens: int | None = None
+    # different pruning behavior is expected when the
+    # user manually selects documents they want to chat with
+    # e.g. we don't want to truncate each document to be no more
+    # than one chunk long
+    is_manually_selected_docs: bool = False
+    # If user specifies to include additional context Chunks for each match, then different pruning
+    # is used. As many Sections as possible are included, and the last Section is truncated
+    # If this is false, all of the Sections are truncated if they are longer than the expected Chunk size.
+    # Sections are often expected to be longer than the maximum Chunk size but Chunks should not be.
+    use_sections: bool = True
+    # If using tools, then we need to consider the tool length
+    tool_num_tokens: int = 0
+    # If using a tool message to represent the docs, then we have to JSON serialize
+    # the document content, which adds to the token count.
+    using_tool_message: bool = False
+
+
+class ContextualPruningConfig(DocumentPruningConfig):
+    num_chunk_multiple: int
+
+    @classmethod
+    def from_doc_pruning_config(
+        cls, num_chunk_multiple: int, doc_pruning_config: DocumentPruningConfig
+    ) -> "ContextualPruningConfig":
+        return cls(num_chunk_multiple=num_chunk_multiple, **doc_pruning_config.dict())
+
+
+class CitationConfig(BaseModel):
+    all_docs_useful: bool = False
+
+
+class QuotesConfig(BaseModel):
+    pass
+
+
+class AnswerStyleConfig(BaseModel):
+    citation_config: CitationConfig | None = None
+    quotes_config: QuotesConfig | None = None
+    document_pruning_config: DocumentPruningConfig = Field(
+        default_factory=DocumentPruningConfig
+    )
+    # forces the LLM to return a structured response, see
+    # https://platform.openai.com/docs/guides/structured-outputs/introduction
+    # right now, only used by the simple chat API
+    structured_response_format: dict | None = None
+
+    @model_validator(mode="after")
+    def check_quotes_and_citation(self) -> "AnswerStyleConfig":
+        if self.citation_config is None and self.quotes_config is None:
+            raise ValueError(
+                "One of `citation_config` or `quotes_config` must be provided"
+            )
+
+        if self.citation_config is not None and self.quotes_config is not None:
+            raise ValueError(
+                "Only one of `citation_config` or `quotes_config` must be provided"
+            )
+
+        return self
+
+
+class PromptConfig(BaseModel):
+    """Final representation of the Prompt configuration passed
+    into the `Answer` object."""
+
+    system_prompt: str
+    task_prompt: str
+    datetime_aware: bool
+    include_citations: bool
+
+    @classmethod
+    def from_model(
+        cls, model: "Prompt", prompt_override: PromptOverride | None = None
+    ) -> "PromptConfig":
+        override_system_prompt = (
+            prompt_override.system_prompt if prompt_override else None
+        )
+        override_task_prompt = prompt_override.task_prompt if prompt_override else None
+
+        return cls(
+            system_prompt=override_system_prompt or model.system_prompt,
+            task_prompt=override_task_prompt or model.task_prompt,
+            datetime_aware=model.datetime_aware,
+            include_citations=model.include_citations,
+        )
+
+    model_config = ConfigDict(frozen=True)
+
+
+ResponsePart = (
+    DanswerAnswerPiece
+    | CitationInfo
+    | ToolCallKickoff
+    | ToolResponse
+    | ToolCallFinalResult
+    | StreamStopInfo
+)

backend/danswer/chat/process_message.py

@@ -6,19 +6,24 @@ from typing import cast
 
 from sqlalchemy.orm import Session
 
+from danswer.chat.answer import Answer
 from danswer.chat.chat_utils import create_chat_chain
 from danswer.chat.chat_utils import create_temporary_persona
 from danswer.chat.models import AllCitations
+from danswer.chat.models import AnswerStyleConfig
 from danswer.chat.models import ChatDanswerBotResponse
+from danswer.chat.models import CitationConfig
 from danswer.chat.models import CitationInfo
 from danswer.chat.models import CustomToolResponse
 from danswer.chat.models import DanswerAnswerPiece
 from danswer.chat.models import DanswerContexts
+from danswer.chat.models import DocumentPruningConfig
 from danswer.chat.models import FileChatDisplay
 from danswer.chat.models import FinalUsedContextDocsResponse
 from danswer.chat.models import LLMRelevanceFilterResponse
 from danswer.chat.models import MessageResponseIDInfo
 from danswer.chat.models import MessageSpecificCitations
+from danswer.chat.models import PromptConfig
 from danswer.chat.models import QADocsResponse
 from danswer.chat.models import StreamingError
 from danswer.chat.models import StreamStopInfo
@@ -57,16 +62,11 @@ from danswer.document_index.factory import get_default_document_index
 from danswer.file_store.models import ChatFileType
 from danswer.file_store.models import FileDescriptor
 from danswer.file_store.utils import load_all_chat_files
-from danswer.file_store.utils import save_files_from_urls
-from danswer.llm.answering.answer import Answer
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import CitationConfig
-from danswer.llm.answering.models import DocumentPruningConfig
-from danswer.llm.answering.models import PreviousMessage
-from danswer.llm.answering.models import PromptConfig
+from danswer.file_store.utils import save_files
 from danswer.llm.exceptions import GenAIDisabledException
 from danswer.llm.factory import get_llms_for_persona
 from danswer.llm.factory import get_main_llm_from_tuple
+from danswer.llm.models import PreviousMessage
 from danswer.llm.utils import litellm_exception_to_error_msg
 from danswer.natural_language_processing.utils import get_tokenizer
 from danswer.server.query_and_chat.models import ChatMessageDetail
@@ -119,6 +119,7 @@ from danswer.utils.logger import setup_logger
 from danswer.utils.long_term_log import LongTermLogger
 from danswer.utils.timing import log_function_time
 from danswer.utils.timing import log_generator_function_time
+from shared_configs.contextvars import CURRENT_TENANT_ID_CONTEXTVAR
 
 
 logger = setup_logger()
@@ -302,6 +303,7 @@ def stream_chat_message_objects(
     3. [always] A set of streamed LLM tokens or an error anywhere along the line if something fails
     4. [always] Details on the final AI response message that is created
     """
+    tenant_id = CURRENT_TENANT_ID_CONTEXTVAR.get()
     use_existing_user_message = new_msg_req.use_existing_user_message
     existing_assistant_message_id = new_msg_req.existing_assistant_message_id
 
@@ -678,7 +680,8 @@ def stream_chat_message_objects(
 
         reference_db_search_docs = None
         qa_docs_response = None
-        ai_message_files = None  # any files to associate with the AI message e.g. dall-e generated images
+        # any files to associate with the AI message e.g. dall-e generated images
+        ai_message_files = []
         dropped_indices = None
         tool_result = None
 
@@ -733,8 +736,14 @@ def stream_chat_message_objects(
                         list[ImageGenerationResponse], packet.response
                     )
 
-                    file_ids = save_files_from_urls(
-                        [img.url for img in img_generation_response]
+                    file_ids = save_files(
+                        urls=[img.url for img in img_generation_response if img.url],
+                        base64_files=[
+                            img.image_data
+                            for img in img_generation_response
+                            if img.image_data
+                        ],
+                        tenant_id=tenant_id,
                     )
                     ai_message_files = [
                         FileDescriptor(id=str(file_id), type=ChatFileType.IMAGE)
@@ -760,15 +769,19 @@ def stream_chat_message_objects(
                         or custom_tool_response.response_type == "csv"
                     ):
                         file_ids = custom_tool_response.tool_result.file_ids
-                        ai_message_files = [
-                            FileDescriptor(
-                                id=str(file_id),
-                                type=ChatFileType.IMAGE
-                                if custom_tool_response.response_type == "image"
-                                else ChatFileType.CSV,
-                            )
-                            for file_id in file_ids
-                        ]
+                        ai_message_files.extend(
+                            [
+                                FileDescriptor(
+                                    id=str(file_id),
+                                    type=(
+                                        ChatFileType.IMAGE
+                                        if custom_tool_response.response_type == "image"
+                                        else ChatFileType.CSV
+                                    ),
+                                )
+                                for file_id in file_ids
+                            ]
+                        )
                         yield FileChatDisplay(
                             file_ids=[str(file_id) for file_id in file_ids]
                         )
@@ -818,7 +831,8 @@ def stream_chat_message_objects(
                 citations_list=answer.citations,
                 db_docs=reference_db_search_docs,
             )
-            yield AllCitations(citations=answer.citations)
+            if not answer.is_cancelled():
+                yield AllCitations(citations=answer.citations)
 
         # Saving Gen AI answer and responding with message info
         tool_name_to_tool_id: dict[str, int] = {}

backend/danswer/chat/prompt_builder/build.py

@@ -4,20 +4,26 @@ from typing import cast
 from langchain_core.messages import BaseMessage
 from langchain_core.messages import HumanMessage
 from langchain_core.messages import SystemMessage
+from pydantic.v1 import BaseModel as BaseModel__v1
 
+from danswer.chat.models import PromptConfig
+from danswer.chat.prompt_builder.citations_prompt import compute_max_llm_input_tokens
+from danswer.chat.prompt_builder.utils import translate_history_to_basemessages
 from danswer.file_store.models import InMemoryChatFile
-from danswer.llm.answering.models import PreviousMessage
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prompts.citations_prompt import compute_max_llm_input_tokens
 from danswer.llm.interfaces import LLMConfig
+from danswer.llm.models import PreviousMessage
 from danswer.llm.utils import build_content_with_imgs
 from danswer.llm.utils import check_message_tokens
 from danswer.llm.utils import message_to_prompt_and_imgs
-from danswer.llm.utils import translate_history_to_basemessages
 from danswer.natural_language_processing.utils import get_tokenizer
 from danswer.prompts.chat_prompts import CHAT_USER_CONTEXT_FREE_PROMPT
 from danswer.prompts.prompt_utils import add_date_time_to_prompt
 from danswer.prompts.prompt_utils import drop_messages_history_overflow
+from danswer.tools.force import ForceUseTool
+from danswer.tools.models import ToolCallFinalResult
+from danswer.tools.models import ToolCallKickoff
+from danswer.tools.models import ToolResponse
+from danswer.tools.tool import Tool
 
 
 def default_build_system_message(
@@ -139,3 +145,15 @@ class AnswerPromptBuilder:
         return drop_messages_history_overflow(
             final_messages_with_tokens, self.max_tokens
         )
+
+
+class LLMCall(BaseModel__v1):
+    prompt_builder: AnswerPromptBuilder
+    tools: list[Tool]
+    force_use_tool: ForceUseTool
+    files: list[InMemoryChatFile]
+    tool_call_info: list[ToolCallKickoff | ToolResponse | ToolCallFinalResult]
+    using_tool_calling_llm: bool
+
+    class Config:
+        arbitrary_types_allowed = True

backend/danswer/chat/prompt_builder/citations_prompt.py

@@ -2,12 +2,12 @@ from langchain.schema.messages import HumanMessage
 from langchain.schema.messages import SystemMessage
 
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import PromptConfig
 from danswer.configs.model_configs import GEN_AI_SINGLE_USER_MESSAGE_EXPECTED_MAX_TOKENS
 from danswer.context.search.models import InferenceChunk
 from danswer.db.models import Persona
 from danswer.db.persona import get_default_prompt__read_only
 from danswer.db.search_settings import get_multilingual_expansion
-from danswer.llm.answering.models import PromptConfig
 from danswer.llm.factory import get_llms_for_persona
 from danswer.llm.factory import get_main_llm_from_tuple
 from danswer.llm.interfaces import LLMConfig

backend/danswer/chat/prompt_builder/quotes_prompt.py

@@ -1,10 +1,10 @@
 from langchain.schema.messages import HumanMessage
 
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import PromptConfig
 from danswer.configs.chat_configs import LANGUAGE_HINT
 from danswer.context.search.models import InferenceChunk
 from danswer.db.search_settings import get_multilingual_expansion
-from danswer.llm.answering.models import PromptConfig
 from danswer.llm.utils import message_to_prompt_and_imgs
 from danswer.prompts.direct_qa_prompts import CONTEXT_BLOCK
 from danswer.prompts.direct_qa_prompts import HISTORY_BLOCK

backend/danswer/chat/prompt_builder/utils.py

@@ -0,0 +1,62 @@
+from langchain.schema.messages import AIMessage
+from langchain.schema.messages import BaseMessage
+from langchain.schema.messages import HumanMessage
+
+from danswer.configs.constants import MessageType
+from danswer.db.models import ChatMessage
+from danswer.file_store.models import InMemoryChatFile
+from danswer.llm.models import PreviousMessage
+from danswer.llm.utils import build_content_with_imgs
+from danswer.prompts.direct_qa_prompts import PARAMATERIZED_PROMPT
+from danswer.prompts.direct_qa_prompts import PARAMATERIZED_PROMPT_WITHOUT_CONTEXT
+
+
+def build_dummy_prompt(
+    system_prompt: str, task_prompt: str, retrieval_disabled: bool
+) -> str:
+    if retrieval_disabled:
+        return PARAMATERIZED_PROMPT_WITHOUT_CONTEXT.format(
+            user_query="<USER_QUERY>",
+            system_prompt=system_prompt,
+            task_prompt=task_prompt,
+        ).strip()
+
+    return PARAMATERIZED_PROMPT.format(
+        context_docs_str="<CONTEXT_DOCS>",
+        user_query="<USER_QUERY>",
+        system_prompt=system_prompt,
+        task_prompt=task_prompt,
+    ).strip()
+
+
+def translate_danswer_msg_to_langchain(
+    msg: ChatMessage | PreviousMessage,
+) -> BaseMessage:
+    files: list[InMemoryChatFile] = []
+
+    # If the message is a `ChatMessage`, it doesn't have the downloaded files
+    # attached. Just ignore them for now.
+    if not isinstance(msg, ChatMessage):
+        files = msg.files
+    content = build_content_with_imgs(msg.message, files, message_type=msg.message_type)
+
+    if msg.message_type == MessageType.SYSTEM:
+        raise ValueError("System messages are not currently part of history")
+    if msg.message_type == MessageType.ASSISTANT:
+        return AIMessage(content=content)
+    if msg.message_type == MessageType.USER:
+        return HumanMessage(content=content)
+
+    raise ValueError(f"New message type {msg.message_type} not handled")
+
+
+def translate_history_to_basemessages(
+    history: list[ChatMessage] | list["PreviousMessage"],
+) -> tuple[list[BaseMessage], list[int]]:
+    history_basemessages = [
+        translate_danswer_msg_to_langchain(msg)
+        for msg in history
+        if msg.token_count != 0
+    ]
+    history_token_counts = [msg.token_count for msg in history if msg.token_count != 0]
+    return history_basemessages, history_token_counts

backend/danswer/chat/prune_and_merge.py

@@ -5,16 +5,16 @@ from typing import TypeVar
 
 from pydantic import BaseModel
 
+from danswer.chat.models import ContextualPruningConfig
 from danswer.chat.models import (
     LlmDoc,
 )
+from danswer.chat.models import PromptConfig
+from danswer.chat.prompt_builder.citations_prompt import compute_max_document_tokens
 from danswer.configs.constants import IGNORE_FOR_QA
 from danswer.configs.model_configs import DOC_EMBEDDING_CONTEXT_SIZE
 from danswer.context.search.models import InferenceChunk
 from danswer.context.search.models import InferenceSection
-from danswer.llm.answering.models import ContextualPruningConfig
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prompts.citations_prompt import compute_max_document_tokens
 from danswer.llm.interfaces import LLMConfig
 from danswer.natural_language_processing.utils import get_tokenizer
 from danswer.natural_language_processing.utils import tokenizer_trim_content

backend/danswer/chat/stream_processing/answer_response_handler.py

@@ -3,13 +3,11 @@ from collections.abc import Generator
 
 from langchain_core.messages import BaseMessage
 
+from danswer.chat.llm_response_handler import ResponsePart
 from danswer.chat.models import CitationInfo
 from danswer.chat.models import LlmDoc
-from danswer.llm.answering.llm_response_handler import ResponsePart
-from danswer.llm.answering.stream_processing.citation_processing import (
-    CitationProcessor,
-)
-from danswer.llm.answering.stream_processing.utils import DocumentIdOrderMapping
+from danswer.chat.stream_processing.citation_processing import CitationProcessor
+from danswer.chat.stream_processing.utils import DocumentIdOrderMapping
 from danswer.utils.logger import setup_logger
 
 logger = setup_logger()

backend/danswer/chat/stream_processing/citation_processing.py

@@ -4,8 +4,8 @@ from collections.abc import Generator
 from danswer.chat.models import CitationInfo
 from danswer.chat.models import DanswerAnswerPiece
 from danswer.chat.models import LlmDoc
+from danswer.chat.stream_processing.utils import DocumentIdOrderMapping
 from danswer.configs.chat_configs import STOP_STREAM_PAT
-from danswer.llm.answering.stream_processing.utils import DocumentIdOrderMapping
 from danswer.prompts.constants import TRIPLE_BACKTICK
 from danswer.utils.logger import setup_logger
 

backend/danswer/chat/tool_handling/tool_response_handler.py

@@ -4,8 +4,8 @@ from langchain_core.messages import AIMessageChunk
 from langchain_core.messages import BaseMessage
 from langchain_core.messages import ToolCall
 
-from danswer.llm.answering.llm_response_handler import LLMCall
-from danswer.llm.answering.llm_response_handler import ResponsePart
+from danswer.chat.models import ResponsePart
+from danswer.chat.prompt_builder.build import LLMCall
 from danswer.llm.interfaces import LLM
 from danswer.tools.force import ForceUseTool
 from danswer.tools.message import build_tool_message

backend/danswer/configs/app_configs.py

@@ -43,9 +43,6 @@ WEB_DOMAIN = os.environ.get("WEB_DOMAIN") or "http://localhost:3000"
 AUTH_TYPE = AuthType((os.environ.get("AUTH_TYPE") or AuthType.DISABLED.value).lower())
 DISABLE_AUTH = AUTH_TYPE == AuthType.DISABLED
 
-# Necessary for cloud integration tests
-DISABLE_VERIFICATION = os.environ.get("DISABLE_VERIFICATION", "").lower() == "true"
-
 # Encryption key secret is used to encrypt connector credentials, api keys, and other sensitive
 # information. This provides an extra layer of security on top of Postgres access controls
 # and is available in Danswer EE
@@ -85,6 +82,7 @@ OAUTH_CLIENT_SECRET = (
 )
 
 USER_AUTH_SECRET = os.environ.get("USER_AUTH_SECRET", "")
+
 # for basic auth
 REQUIRE_EMAIL_VERIFICATION = (
     os.environ.get("REQUIRE_EMAIL_VERIFICATION", "").lower() == "true"
@@ -118,6 +116,8 @@ VESPA_HOST = os.environ.get("VESPA_HOST") or "localhost"
 VESPA_CONFIG_SERVER_HOST = os.environ.get("VESPA_CONFIG_SERVER_HOST") or VESPA_HOST
 VESPA_PORT = os.environ.get("VESPA_PORT") or "8081"
 VESPA_TENANT_PORT = os.environ.get("VESPA_TENANT_PORT") or "19071"
+# the number of times to try and connect to vespa on startup before giving up
+VESPA_NUM_ATTEMPTS_ON_STARTUP = int(os.environ.get("NUM_RETRIES_ON_STARTUP") or 10)
 
 VESPA_CLOUD_URL = os.environ.get("VESPA_CLOUD_URL", "")
 

backend/danswer/configs/tool_configs.py

@@ -2,6 +2,8 @@ import json
 import os
 
 
+IMAGE_GENERATION_OUTPUT_FORMAT = os.environ.get("IMAGE_GENERATION_OUTPUT_FORMAT", "url")
+
 # if specified, will pass through request headers to the call to API calls made by custom tools
 CUSTOM_TOOL_PASS_THROUGH_HEADERS: list[str] | None = None
 _CUSTOM_TOOL_PASS_THROUGH_HEADERS_RAW = os.environ.get(

backend/danswer/connectors/confluence/connector.py

@@ -15,6 +15,7 @@ from danswer.connectors.confluence.utils import attachment_to_content
 from danswer.connectors.confluence.utils import build_confluence_document_id
 from danswer.connectors.confluence.utils import datetime_from_string
 from danswer.connectors.confluence.utils import extract_text_from_confluence_html
+from danswer.connectors.confluence.utils import validate_attachment_filetype
 from danswer.connectors.interfaces import GenerateDocumentsOutput
 from danswer.connectors.interfaces import GenerateSlimDocumentOutput
 from danswer.connectors.interfaces import LoadConnector
@@ -276,9 +277,11 @@ class ConfluenceConnector(LoadConnector, PollConnector, SlimConnector):
         ):
             # If the page has restrictions, add them to the perm_sync_data
             # These will be used by doc_sync.py to sync permissions
-            perm_sync_data = {
-                "restrictions": page.get("restrictions", {}),
-                "space_key": page.get("space", {}).get("key"),
+            page_restrictions = page.get("restrictions")
+            page_space_key = page.get("space", {}).get("key")
+            page_perm_sync_data = {
+                "restrictions": page_restrictions or {},
+                "space_key": page_space_key,
             }
 
             doc_metadata_list.append(
@@ -288,7 +291,7 @@ class ConfluenceConnector(LoadConnector, PollConnector, SlimConnector):
                         page["_links"]["webui"],
                         self.is_cloud,
                     ),
-                    perm_sync_data=perm_sync_data,
+                    perm_sync_data=page_perm_sync_data,
                 )
             )
             attachment_cql = f"type=attachment and container='{page['id']}'"
@@ -298,6 +301,21 @@ class ConfluenceConnector(LoadConnector, PollConnector, SlimConnector):
                 expand=restrictions_expand,
                 limit=_SLIM_DOC_BATCH_SIZE,
             ):
+                if not validate_attachment_filetype(attachment):
+                    continue
+                attachment_restrictions = attachment.get("restrictions")
+                if not attachment_restrictions:
+                    attachment_restrictions = page_restrictions
+
+                attachment_space_key = attachment.get("space", {}).get("key")
+                if not attachment_space_key:
+                    attachment_space_key = page_space_key
+
+                attachment_perm_sync_data = {
+                    "restrictions": attachment_restrictions or {},
+                    "space_key": attachment_space_key,
+                }
+
                 doc_metadata_list.append(
                     SlimDocument(
                         id=build_confluence_document_id(
@@ -305,7 +323,7 @@ class ConfluenceConnector(LoadConnector, PollConnector, SlimConnector):
                             attachment["_links"]["webui"],
                             self.is_cloud,
                         ),
-                        perm_sync_data=perm_sync_data,
+                        perm_sync_data=attachment_perm_sync_data,
                     )
                 )
             if len(doc_metadata_list) > _SLIM_DOC_BATCH_SIZE:

backend/danswer/connectors/confluence/utils.py

@@ -177,19 +177,23 @@ def extract_text_from_confluence_html(
     return format_document_soup(soup)
 
 
-def attachment_to_content(
-    confluence_client: OnyxConfluence,
-    attachment: dict[str, Any],
-) -> str | None:
-    """If it returns None, assume that we should skip this attachment."""
-    if attachment["metadata"]["mediaType"] in [
+def validate_attachment_filetype(attachment: dict[str, Any]) -> bool:
+    return attachment["metadata"]["mediaType"] not in [
         "image/jpeg",
         "image/png",
         "image/gif",
         "image/svg+xml",
         "video/mp4",
         "video/quicktime",
-    ]:
+    ]
+
+
+def attachment_to_content(
+    confluence_client: OnyxConfluence,
+    attachment: dict[str, Any],
+) -> str | None:
+    """If it returns None, assume that we should skip this attachment."""
+    if not validate_attachment_filetype(attachment):
         return None
 
     download_link = confluence_client.url + attachment["_links"]["download"]
@@ -245,7 +249,7 @@ def build_confluence_document_id(
     return f"{base_url}{content_url}"
 
 
-def extract_referenced_attachment_names(page_text: str) -> list[str]:
+def _extract_referenced_attachment_names(page_text: str) -> list[str]:
     """Parse a Confluence html page to generate a list of current
         attachments in use
 

backend/danswer/context/search/pipeline.py

@@ -5,7 +5,11 @@ from typing import cast
 
 from sqlalchemy.orm import Session
 
+from danswer.chat.models import PromptConfig
 from danswer.chat.models import SectionRelevancePiece
+from danswer.chat.prune_and_merge import _merge_sections
+from danswer.chat.prune_and_merge import ChunkRange
+from danswer.chat.prune_and_merge import merge_chunk_intervals
 from danswer.configs.chat_configs import DISABLE_LLM_DOC_RELEVANCE
 from danswer.context.search.enums import LLMEvaluationType
 from danswer.context.search.enums import QueryFlow
@@ -27,10 +31,6 @@ from danswer.db.models import User
 from danswer.db.search_settings import get_current_search_settings
 from danswer.document_index.factory import get_default_document_index
 from danswer.document_index.interfaces import VespaChunkRequest
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prune_and_merge import _merge_sections
-from danswer.llm.answering.prune_and_merge import ChunkRange
-from danswer.llm.answering.prune_and_merge import merge_chunk_intervals
 from danswer.llm.interfaces import LLM
 from danswer.secondary_llm_flows.agentic_evaluation import evaluate_inference_section
 from danswer.utils.logger import setup_logger

backend/danswer/document_index/vespa/app_config/schemas/danswer_chunk.sd

@@ -4,6 +4,8 @@ schema DANSWER_CHUNK_NAME {
         # Not to be confused with the UUID generated for this chunk which is called documentid by default
         field document_id type string {
             indexing: summary | attribute
+            attribute: fast-search
+            rank: filter
         }
         field chunk_id type int {
             indexing: summary | attribute

backend/danswer/file_processing/extract_file_text.py

@@ -6,6 +6,7 @@ import zipfile
 from collections.abc import Callable
 from collections.abc import Iterator
 from email.parser import Parser as EmailParser
+from io import BytesIO
 from pathlib import Path
 from typing import Any
 from typing import Dict
@@ -15,13 +16,17 @@ import chardet
 import docx  # type: ignore
 import openpyxl  # type: ignore
 import pptx  # type: ignore
+from docx import Document
+from fastapi import UploadFile
 from pypdf import PdfReader
 from pypdf.errors import PdfStreamError
 
 from danswer.configs.constants import DANSWER_METADATA_FILENAME
+from danswer.configs.constants import FileOrigin
 from danswer.file_processing.html_utils import parse_html_page_basic
 from danswer.file_processing.unstructured import get_unstructured_api_key
 from danswer.file_processing.unstructured import unstructured_to_text
+from danswer.file_store.file_store import FileStore
 from danswer.utils.logger import setup_logger
 
 logger = setup_logger()
@@ -375,3 +380,35 @@ def extract_file_text(
             ) from e
         logger.warning(f"Failed to process file {file_name or 'Unknown'}: {str(e)}")
         return ""
+
+
+def convert_docx_to_txt(
+    file: UploadFile, file_store: FileStore, file_path: str
+) -> None:
+    file.file.seek(0)
+    docx_content = file.file.read()
+    doc = Document(BytesIO(docx_content))
+
+    # Extract text from the document
+    full_text = []
+    for para in doc.paragraphs:
+        full_text.append(para.text)
+
+    # Join the extracted text
+    text_content = "\n".join(full_text)
+
+    txt_file_path = docx_to_txt_filename(file_path)
+    file_store.save_file(
+        file_name=txt_file_path,
+        content=BytesIO(text_content.encode("utf-8")),
+        display_name=file.filename,
+        file_origin=FileOrigin.CONNECTOR,
+        file_type="text/plain",
+    )
+
+
+def docx_to_txt_filename(file_path: str) -> str:
+    """
+    Convert a .docx file path to its corresponding .txt file path.
+    """
+    return file_path.rsplit(".", 1)[0] + ".txt"

backend/danswer/file_store/utils.py

@@ -1,6 +1,6 @@
+import base64
 from collections.abc import Callable
 from io import BytesIO
-from typing import Any
 from typing import cast
 from uuid import uuid4
 
@@ -13,8 +13,8 @@ from danswer.db.models import ChatMessage
 from danswer.file_store.file_store import get_default_file_store
 from danswer.file_store.models import FileDescriptor
 from danswer.file_store.models import InMemoryChatFile
+from danswer.utils.b64 import get_image_type
 from danswer.utils.threadpool_concurrency import run_functions_tuples_in_parallel
-from shared_configs.contextvars import CURRENT_TENANT_ID_CONTEXTVAR
 
 
 def load_chat_file(
@@ -75,11 +75,58 @@ def save_file_from_url(url: str, tenant_id: str) -> str:
         return unique_id
 
 
-def save_files_from_urls(urls: list[str]) -> list[str]:
-    tenant_id = CURRENT_TENANT_ID_CONTEXTVAR.get()
+def save_file_from_base64(base64_string: str, tenant_id: str) -> str:
+    with get_session_with_tenant(tenant_id) as db_session:
+        unique_id = str(uuid4())
+        file_store = get_default_file_store(db_session)
+        file_store.save_file(
+            file_name=unique_id,
+            content=BytesIO(base64.b64decode(base64_string)),
+            display_name="GeneratedImage",
+            file_origin=FileOrigin.CHAT_IMAGE_GEN,
+            file_type=get_image_type(base64_string),
+        )
+        return unique_id
 
-    funcs: list[tuple[Callable[..., Any], tuple[Any, ...]]] = [
-        (save_file_from_url, (url, tenant_id)) for url in urls
+
+def save_file(
+    tenant_id: str,
+    url: str | None = None,
+    base64_data: str | None = None,
+) -> str:
+    """Save a file from either a URL or base64 encoded string.
+
+    Args:
+        tenant_id: The tenant ID to save the file under
+        url: URL to download file from
+        base64_data: Base64 encoded file data
+
+    Returns:
+        The unique ID of the saved file
+
+    Raises:
+        ValueError: If neither url nor base64_data is provided, or if both are provided
+    """
+    if url is not None and base64_data is not None:
+        raise ValueError("Cannot specify both url and base64_data")
+
+    if url is not None:
+        return save_file_from_url(url, tenant_id)
+    elif base64_data is not None:
+        return save_file_from_base64(base64_data, tenant_id)
+    else:
+        raise ValueError("Must specify either url or base64_data")
+
+
+def save_files(urls: list[str], base64_files: list[str], tenant_id: str) -> list[str]:
+    # NOTE: be explicit about typing so that if we change things, we get notified
+    funcs: list[
+        tuple[
+            Callable[[str, str | None, str | None], str],
+            tuple[str, str | None, str | None],
+        ]
+    ] = [(save_file, (tenant_id, url, None)) for url in urls] + [
+        (save_file, (tenant_id, None, base64_file)) for base64_file in base64_files
     ]
-    # Must pass in tenant_id here, since this is called by multithreading
+
     return run_functions_tuples_in_parallel(funcs)

backend/danswer/llm/answering/models.py

@@ -1,163 +0,0 @@
-from collections.abc import Callable
-from collections.abc import Iterator
-from typing import TYPE_CHECKING
-
-from langchain.schema.messages import AIMessage
-from langchain.schema.messages import BaseMessage
-from langchain.schema.messages import HumanMessage
-from langchain.schema.messages import SystemMessage
-from pydantic import BaseModel
-from pydantic import ConfigDict
-from pydantic import Field
-from pydantic import model_validator
-
-from danswer.chat.models import AnswerQuestionStreamReturn
-from danswer.configs.constants import MessageType
-from danswer.file_store.models import InMemoryChatFile
-from danswer.llm.override_models import PromptOverride
-from danswer.llm.utils import build_content_with_imgs
-from danswer.tools.models import ToolCallFinalResult
-
-if TYPE_CHECKING:
-    from danswer.db.models import ChatMessage
-    from danswer.db.models import Prompt
-
-
-StreamProcessor = Callable[[Iterator[str]], AnswerQuestionStreamReturn]
-
-
-class PreviousMessage(BaseModel):
-    """Simplified version of `ChatMessage`"""
-
-    message: str
-    token_count: int
-    message_type: MessageType
-    files: list[InMemoryChatFile]
-    tool_call: ToolCallFinalResult | None
-
-    @classmethod
-    def from_chat_message(
-        cls, chat_message: "ChatMessage", available_files: list[InMemoryChatFile]
-    ) -> "PreviousMessage":
-        message_file_ids = (
-            [file["id"] for file in chat_message.files] if chat_message.files else []
-        )
-        return cls(
-            message=chat_message.message,
-            token_count=chat_message.token_count,
-            message_type=chat_message.message_type,
-            files=[
-                file
-                for file in available_files
-                if str(file.file_id) in message_file_ids
-            ],
-            tool_call=ToolCallFinalResult(
-                tool_name=chat_message.tool_call.tool_name,
-                tool_args=chat_message.tool_call.tool_arguments,
-                tool_result=chat_message.tool_call.tool_result,
-            )
-            if chat_message.tool_call
-            else None,
-        )
-
-    def to_langchain_msg(self) -> BaseMessage:
-        content = build_content_with_imgs(self.message, self.files)
-        if self.message_type == MessageType.USER:
-            return HumanMessage(content=content)
-        elif self.message_type == MessageType.ASSISTANT:
-            return AIMessage(content=content)
-        else:
-            return SystemMessage(content=content)
-
-
-class DocumentPruningConfig(BaseModel):
-    max_chunks: int | None = None
-    max_window_percentage: float | None = None
-    max_tokens: int | None = None
-    # different pruning behavior is expected when the
-    # user manually selects documents they want to chat with
-    # e.g. we don't want to truncate each document to be no more
-    # than one chunk long
-    is_manually_selected_docs: bool = False
-    # If user specifies to include additional context Chunks for each match, then different pruning
-    # is used. As many Sections as possible are included, and the last Section is truncated
-    # If this is false, all of the Sections are truncated if they are longer than the expected Chunk size.
-    # Sections are often expected to be longer than the maximum Chunk size but Chunks should not be.
-    use_sections: bool = True
-    # If using tools, then we need to consider the tool length
-    tool_num_tokens: int = 0
-    # If using a tool message to represent the docs, then we have to JSON serialize
-    # the document content, which adds to the token count.
-    using_tool_message: bool = False
-
-
-class ContextualPruningConfig(DocumentPruningConfig):
-    num_chunk_multiple: int
-
-    @classmethod
-    def from_doc_pruning_config(
-        cls, num_chunk_multiple: int, doc_pruning_config: DocumentPruningConfig
-    ) -> "ContextualPruningConfig":
-        return cls(num_chunk_multiple=num_chunk_multiple, **doc_pruning_config.dict())
-
-
-class CitationConfig(BaseModel):
-    all_docs_useful: bool = False
-
-
-class QuotesConfig(BaseModel):
-    pass
-
-
-class AnswerStyleConfig(BaseModel):
-    citation_config: CitationConfig | None = None
-    quotes_config: QuotesConfig | None = None
-    document_pruning_config: DocumentPruningConfig = Field(
-        default_factory=DocumentPruningConfig
-    )
-    # forces the LLM to return a structured response, see
-    # https://platform.openai.com/docs/guides/structured-outputs/introduction
-    # right now, only used by the simple chat API
-    structured_response_format: dict | None = None
-
-    @model_validator(mode="after")
-    def check_quotes_and_citation(self) -> "AnswerStyleConfig":
-        if self.citation_config is None and self.quotes_config is None:
-            raise ValueError(
-                "One of `citation_config` or `quotes_config` must be provided"
-            )
-
-        if self.citation_config is not None and self.quotes_config is not None:
-            raise ValueError(
-                "Only one of `citation_config` or `quotes_config` must be provided"
-            )
-
-        return self
-
-
-class PromptConfig(BaseModel):
-    """Final representation of the Prompt configuration passed
-    into the `Answer` object."""
-
-    system_prompt: str
-    task_prompt: str
-    datetime_aware: bool
-    include_citations: bool
-
-    @classmethod
-    def from_model(
-        cls, model: "Prompt", prompt_override: PromptOverride | None = None
-    ) -> "PromptConfig":
-        override_system_prompt = (
-            prompt_override.system_prompt if prompt_override else None
-        )
-        override_task_prompt = prompt_override.task_prompt if prompt_override else None
-
-        return cls(
-            system_prompt=override_system_prompt or model.system_prompt,
-            task_prompt=override_task_prompt or model.task_prompt,
-            datetime_aware=model.datetime_aware,
-            include_citations=model.include_citations,
-        )
-
-    model_config = ConfigDict(frozen=True)

backend/danswer/llm/answering/prompts/utils.py

@@ -1,20 +0,0 @@
-from danswer.prompts.direct_qa_prompts import PARAMATERIZED_PROMPT
-from danswer.prompts.direct_qa_prompts import PARAMATERIZED_PROMPT_WITHOUT_CONTEXT
-
-
-def build_dummy_prompt(
-    system_prompt: str, task_prompt: str, retrieval_disabled: bool
-) -> str:
-    if retrieval_disabled:
-        return PARAMATERIZED_PROMPT_WITHOUT_CONTEXT.format(
-            user_query="<USER_QUERY>",
-            system_prompt=system_prompt,
-            task_prompt=task_prompt,
-        ).strip()
-
-    return PARAMATERIZED_PROMPT.format(
-        context_docs_str="<CONTEXT_DOCS>",
-        user_query="<USER_QUERY>",
-        system_prompt=system_prompt,
-        task_prompt=task_prompt,
-    ).strip()

backend/danswer/llm/models.py

@@ -0,0 +1,59 @@
+from typing import TYPE_CHECKING
+
+from langchain.schema.messages import AIMessage
+from langchain.schema.messages import BaseMessage
+from langchain.schema.messages import HumanMessage
+from langchain.schema.messages import SystemMessage
+from pydantic import BaseModel
+
+from danswer.configs.constants import MessageType
+from danswer.file_store.models import InMemoryChatFile
+from danswer.llm.utils import build_content_with_imgs
+from danswer.tools.models import ToolCallFinalResult
+
+if TYPE_CHECKING:
+    from danswer.db.models import ChatMessage
+
+
+class PreviousMessage(BaseModel):
+    """Simplified version of `ChatMessage`"""
+
+    message: str
+    token_count: int
+    message_type: MessageType
+    files: list[InMemoryChatFile]
+    tool_call: ToolCallFinalResult | None
+
+    @classmethod
+    def from_chat_message(
+        cls, chat_message: "ChatMessage", available_files: list[InMemoryChatFile]
+    ) -> "PreviousMessage":
+        message_file_ids = (
+            [file["id"] for file in chat_message.files] if chat_message.files else []
+        )
+        return cls(
+            message=chat_message.message,
+            token_count=chat_message.token_count,
+            message_type=chat_message.message_type,
+            files=[
+                file
+                for file in available_files
+                if str(file.file_id) in message_file_ids
+            ],
+            tool_call=ToolCallFinalResult(
+                tool_name=chat_message.tool_call.tool_name,
+                tool_args=chat_message.tool_call.tool_arguments,
+                tool_result=chat_message.tool_call.tool_result,
+            )
+            if chat_message.tool_call
+            else None,
+        )
+
+    def to_langchain_msg(self) -> BaseMessage:
+        content = build_content_with_imgs(self.message, self.files)
+        if self.message_type == MessageType.USER:
+            return HumanMessage(content=content)
+        elif self.message_type == MessageType.ASSISTANT:
+            return AIMessage(content=content)
+        else:
+            return SystemMessage(content=content)

backend/danswer/llm/utils.py

@@ -5,8 +5,6 @@ from collections.abc import Callable
 from collections.abc import Iterator
 from typing import Any
 from typing import cast
-from typing import TYPE_CHECKING
-from typing import Union
 
 import litellm  # type: ignore
 import pandas as pd
@@ -36,17 +34,15 @@ from danswer.configs.constants import MessageType
 from danswer.configs.model_configs import GEN_AI_MAX_TOKENS
 from danswer.configs.model_configs import GEN_AI_MODEL_FALLBACK_MAX_TOKENS
 from danswer.configs.model_configs import GEN_AI_NUM_RESERVED_OUTPUT_TOKENS
-from danswer.db.models import ChatMessage
 from danswer.file_store.models import ChatFileType
 from danswer.file_store.models import InMemoryChatFile
 from danswer.llm.interfaces import LLM
 from danswer.prompts.constants import CODE_BLOCK_PAT
+from danswer.utils.b64 import get_image_type
+from danswer.utils.b64 import get_image_type_from_bytes
 from danswer.utils.logger import setup_logger
 from shared_configs.configs import LOG_LEVEL
 
-if TYPE_CHECKING:
-    from danswer.llm.answering.models import PreviousMessage
-
 logger = setup_logger()
 
 
@@ -104,39 +100,6 @@ def litellm_exception_to_error_msg(
     return error_msg
 
 
-def translate_danswer_msg_to_langchain(
-    msg: Union[ChatMessage, "PreviousMessage"],
-) -> BaseMessage:
-    files: list[InMemoryChatFile] = []
-
-    # If the message is a `ChatMessage`, it doesn't have the downloaded files
-    # attached. Just ignore them for now.
-    if not isinstance(msg, ChatMessage):
-        files = msg.files
-    content = build_content_with_imgs(msg.message, files, message_type=msg.message_type)
-
-    if msg.message_type == MessageType.SYSTEM:
-        raise ValueError("System messages are not currently part of history")
-    if msg.message_type == MessageType.ASSISTANT:
-        return AIMessage(content=content)
-    if msg.message_type == MessageType.USER:
-        return HumanMessage(content=content)
-
-    raise ValueError(f"New message type {msg.message_type} not handled")
-
-
-def translate_history_to_basemessages(
-    history: list[ChatMessage] | list["PreviousMessage"],
-) -> tuple[list[BaseMessage], list[int]]:
-    history_basemessages = [
-        translate_danswer_msg_to_langchain(msg)
-        for msg in history
-        if msg.token_count != 0
-    ]
-    history_token_counts = [msg.token_count for msg in history if msg.token_count != 0]
-    return history_basemessages, history_token_counts
-
-
 #  Processes CSV files to show the first 5 rows and max_columns (default 40) columns
 def _process_csv_file(file: InMemoryChatFile, max_columns: int = 40) -> str:
     df = pd.read_csv(io.StringIO(file.content.decode("utf-8")))
@@ -190,6 +153,7 @@ def build_content_with_imgs(
     message: str,
     files: list[InMemoryChatFile] | None = None,
     img_urls: list[str] | None = None,
+    b64_imgs: list[str] | None = None,
     message_type: MessageType = MessageType.USER,
 ) -> str | list[str | dict[str, Any]]:  # matching Langchain's BaseMessage content type
     files = files or []
@@ -202,6 +166,7 @@ def build_content_with_imgs(
     )
 
     img_urls = img_urls or []
+    b64_imgs = b64_imgs or []
 
     message_main_content = _build_content(message, files)
 
@@ -220,11 +185,22 @@ def build_content_with_imgs(
             {
                 "type": "image_url",
                 "image_url": {
-                    "url": f"data:image/jpeg;base64,{file.to_base64()}",
+                    "url": (
+                        f"data:{get_image_type_from_bytes(file.content)};"
+                        f"base64,{file.to_base64()}"
+                    ),
                 },
             }
-            for file in files
-            if file.file_type == "image"
+            for file in img_files
+        ]
+        + [
+            {
+                "type": "image_url",
+                "image_url": {
+                    "url": f"data:{get_image_type(b64_img)};base64,{b64_img}",
+                },
+            }
+            for b64_img in b64_imgs
         ]
         + [
             {

backend/danswer/prompts/prompt_utils.py

@@ -5,11 +5,11 @@ from typing import cast
 from langchain_core.messages import BaseMessage
 
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import PromptConfig
 from danswer.configs.chat_configs import LANGUAGE_HINT
 from danswer.configs.constants import DocumentSource
 from danswer.context.search.models import InferenceChunk
 from danswer.db.models import Prompt
-from danswer.llm.answering.models import PromptConfig
 from danswer.prompts.chat_prompts import ADDITIONAL_INFO
 from danswer.prompts.chat_prompts import CITATION_REMINDER
 from danswer.prompts.constants import CODE_BLOCK_PAT

backend/danswer/secondary_llm_flows/choose_search.py

@@ -3,14 +3,14 @@ from langchain.schema import HumanMessage
 from langchain.schema import SystemMessage
 
 from danswer.chat.chat_utils import combine_message_chain
+from danswer.chat.prompt_builder.utils import translate_danswer_msg_to_langchain
 from danswer.configs.chat_configs import DISABLE_LLM_CHOOSE_SEARCH
 from danswer.configs.model_configs import GEN_AI_HISTORY_CUTOFF
 from danswer.db.models import ChatMessage
-from danswer.llm.answering.models import PreviousMessage
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.llm.utils import dict_based_prompt_to_langchain_prompt
 from danswer.llm.utils import message_to_string
-from danswer.llm.utils import translate_danswer_msg_to_langchain
 from danswer.prompts.chat_prompts import AGGRESSIVE_SEARCH_TEMPLATE
 from danswer.prompts.chat_prompts import NO_SEARCH
 from danswer.prompts.chat_prompts import REQUIRE_SEARCH_HINT

backend/danswer/secondary_llm_flows/query_expansion.py

@@ -4,10 +4,10 @@ from danswer.chat.chat_utils import combine_message_chain
 from danswer.configs.chat_configs import DISABLE_LLM_QUERY_REPHRASE
 from danswer.configs.model_configs import GEN_AI_HISTORY_CUTOFF
 from danswer.db.models import ChatMessage
-from danswer.llm.answering.models import PreviousMessage
 from danswer.llm.exceptions import GenAIDisabledException
 from danswer.llm.factory import get_default_llms
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.llm.utils import dict_based_prompt_to_langchain_prompt
 from danswer.llm.utils import message_to_string
 from danswer.prompts.chat_prompts import HISTORY_QUERY_REPHRASE

backend/danswer/server/documents/connector.py

@@ -86,6 +86,7 @@ from danswer.db.models import SearchSettings
 from danswer.db.models import User
 from danswer.db.search_settings import get_current_search_settings
 from danswer.db.search_settings import get_secondary_search_settings
+from danswer.file_processing.extract_file_text import convert_docx_to_txt
 from danswer.file_store.file_store import get_default_file_store
 from danswer.key_value_store.interface import KvKeyNotFoundError
 from danswer.redis.redis_connector import RedisConnector
@@ -393,6 +394,12 @@ def upload_files(
                 file_origin=FileOrigin.CONNECTOR,
                 file_type=file.content_type or "text/plain",
             )
+
+            if file.content_type and file.content_type.startswith(
+                "application/vnd.openxmlformats-officedocument.wordprocessingml.document"
+            ):
+                convert_docx_to_txt(file, file_store, file_path)
+
     except ValueError as e:
         raise HTTPException(status_code=400, detail=str(e))
     return FileUploadResponse(file_paths=deduped_file_paths)
@@ -1010,37 +1017,18 @@ def get_connector_by_id(
 
 
 class BasicCCPairInfo(BaseModel):
-    docs_indexed: int
     has_successful_run: bool
     source: DocumentSource
 
 
-@router.get("/indexing-status")
+@router.get("/connector-status")
 def get_basic_connector_indexing_status(
     _: User = Depends(current_user),
     db_session: Session = Depends(get_session),
 ) -> list[BasicCCPairInfo]:
     cc_pairs = get_connector_credential_pairs(db_session)
-    cc_pair_identifiers = [
-        ConnectorCredentialPairIdentifier(
-            connector_id=cc_pair.connector_id, credential_id=cc_pair.credential_id
-        )
-        for cc_pair in cc_pairs
-    ]
-    document_count_info = get_document_counts_for_cc_pairs(
-        db_session=db_session,
-        cc_pair_identifiers=cc_pair_identifiers,
-    )
-    cc_pair_to_document_cnt = {
-        (connector_id, credential_id): cnt
-        for connector_id, credential_id, cnt in document_count_info
-    }
     return [
         BasicCCPairInfo(
-            docs_indexed=cc_pair_to_document_cnt.get(
-                (cc_pair.connector_id, cc_pair.credential_id)
-            )
-            or 0,
             has_successful_run=cc_pair.last_successful_index_time is not None,
             source=cc_pair.connector.source,
         )

backend/danswer/server/features/persona/api.py

@@ -13,6 +13,7 @@ from danswer.auth.users import current_admin_user
 from danswer.auth.users import current_curator_or_admin_user
 from danswer.auth.users import current_limited_user
 from danswer.auth.users import current_user
+from danswer.chat.prompt_builder.utils import build_dummy_prompt
 from danswer.configs.constants import FileOrigin
 from danswer.configs.constants import NotificationType
 from danswer.db.engine import get_session
@@ -33,7 +34,6 @@ from danswer.db.persona import update_persona_shared_users
 from danswer.db.persona import update_persona_visibility
 from danswer.file_store.file_store import get_default_file_store
 from danswer.file_store.models import ChatFileType
-from danswer.llm.answering.prompts.utils import build_dummy_prompt
 from danswer.server.features.persona.models import CreatePersonaRequest
 from danswer.server.features.persona.models import ImageGenerationToolStatus
 from danswer.server.features.persona.models import PersonaCategoryCreate

backend/danswer/server/manage/users.py

@@ -194,11 +194,11 @@ def bulk_invite_users(
         )
 
     tenant_id = CURRENT_TENANT_ID_CONTEXTVAR.get()
-    normalized_emails = []
+    new_invited_emails = []
     try:
         for email in emails:
             email_info = validate_email(email)
-            normalized_emails.append(email_info.normalized)  # type: ignore
+            new_invited_emails.append(email_info.normalized)
 
     except (EmailUndeliverableError, EmailNotValidError) as e:
         raise HTTPException(
@@ -210,7 +210,7 @@ def bulk_invite_users(
         try:
             fetch_ee_implementation_or_noop(
                 "danswer.server.tenants.provisioning", "add_users_to_tenant", None
-            )(normalized_emails, tenant_id)
+            )(new_invited_emails, tenant_id)
 
         except IntegrityError as e:
             if isinstance(e.orig, UniqueViolation):
@@ -224,7 +224,7 @@ def bulk_invite_users(
 
     initial_invited_users = get_invited_users()
 
-    all_emails = list(set(normalized_emails) | set(initial_invited_users))
+    all_emails = list(set(new_invited_emails) | set(initial_invited_users))
     number_of_invited_users = write_invited_users(all_emails)
 
     if not MULTI_TENANT:
@@ -236,7 +236,7 @@ def bulk_invite_users(
         )(CURRENT_TENANT_ID_CONTEXTVAR.get(), get_total_users_count(db_session))
         if ENABLE_EMAIL_INVITES:
             try:
-                for email in all_emails:
+                for email in new_invited_emails:
                     send_user_email_invite(email, current_user)
             except Exception as e:
                 logger.error(f"Error sending email invite to invited users: {e}")
@@ -250,7 +250,7 @@ def bulk_invite_users(
         write_invited_users(initial_invited_users)  # Reset to original state
         fetch_ee_implementation_or_noop(
             "danswer.server.tenants.user_mapping", "remove_users_from_tenant", None
-        )(normalized_emails, tenant_id)
+        )(new_invited_emails, tenant_id)
         raise e
 
 

backend/danswer/server/query_and_chat/chat_backend.py

@@ -1,6 +1,7 @@
 import asyncio
 import io
 import json
+import os
 import uuid
 from collections.abc import Callable
 from collections.abc import Generator
@@ -23,6 +24,9 @@ from danswer.auth.users import current_user
 from danswer.chat.chat_utils import create_chat_chain
 from danswer.chat.chat_utils import extract_headers
 from danswer.chat.process_message import stream_chat_message
+from danswer.chat.prompt_builder.citations_prompt import (
+    compute_max_document_tokens_for_persona,
+)
 from danswer.configs.app_configs import WEB_DOMAIN
 from danswer.configs.constants import FileOrigin
 from danswer.configs.constants import MessageType
@@ -47,13 +51,11 @@ from danswer.db.models import User
 from danswer.db.persona import get_persona_by_id
 from danswer.document_index.document_index_utils import get_both_index_names
 from danswer.document_index.factory import get_default_document_index
+from danswer.file_processing.extract_file_text import docx_to_txt_filename
 from danswer.file_processing.extract_file_text import extract_file_text
 from danswer.file_store.file_store import get_default_file_store
 from danswer.file_store.models import ChatFileType
 from danswer.file_store.models import FileDescriptor
-from danswer.llm.answering.prompts.citations_prompt import (
-    compute_max_document_tokens_for_persona,
-)
 from danswer.llm.exceptions import GenAIDisabledException
 from danswer.llm.factory import get_default_llms
 from danswer.llm.factory import get_llms_for_persona
@@ -718,6 +720,18 @@ def fetch_chat_file(
     if not file_record:
         raise HTTPException(status_code=404, detail="File not found")
 
+    original_file_name = file_record.display_name
+    if file_record.file_type.startswith(
+        "application/vnd.openxmlformats-officedocument.wordprocessingml.document"
+    ):
+        # Check if a converted text file exists for .docx files
+        txt_file_name = docx_to_txt_filename(original_file_name)
+        txt_file_id = os.path.join(os.path.dirname(file_id), txt_file_name)
+        txt_file_record = file_store.read_file_record(txt_file_id)
+        if txt_file_record:
+            file_record = txt_file_record
+            file_id = txt_file_id
+
     media_type = file_record.file_type
     file_io = file_store.read_file(file_id, mode="b")
 

backend/danswer/server/query_and_chat/models.py

@@ -1,5 +1,6 @@
 from datetime import datetime
 from typing import Any
+from typing import TYPE_CHECKING
 from uuid import UUID
 
 from pydantic import BaseModel
@@ -22,6 +23,9 @@ from danswer.llm.override_models import LLMOverride
 from danswer.llm.override_models import PromptOverride
 from danswer.tools.models import ToolCallFinalResult
 
+if TYPE_CHECKING:
+    pass
+
 
 class SourceTag(Tag):
     source: DocumentSource

backend/danswer/setup.py

@@ -4,6 +4,7 @@ from sqlalchemy.orm import Session
 
 from danswer.configs.app_configs import DISABLE_INDEX_UPDATE_ON_SWAP
 from danswer.configs.app_configs import MANAGED_VESPA
+from danswer.configs.app_configs import VESPA_NUM_ATTEMPTS_ON_STARTUP
 from danswer.configs.constants import KV_REINDEX_KEY
 from danswer.configs.constants import KV_SEARCH_SETTINGS
 from danswer.configs.model_configs import FAST_GEN_AI_MODEL_VERSION
@@ -221,13 +222,13 @@ def setup_vespa(
     document_index: DocumentIndex,
     index_setting: IndexingSetting,
     secondary_index_setting: IndexingSetting | None,
+    num_attempts: int = VESPA_NUM_ATTEMPTS_ON_STARTUP,
 ) -> bool:
     # Vespa startup is a bit slow, so give it a few seconds
     WAIT_SECONDS = 5
-    VESPA_ATTEMPTS = 5
-    for x in range(VESPA_ATTEMPTS):
+    for x in range(num_attempts):
         try:
-            logger.notice(f"Setting up Vespa (attempt {x+1}/{VESPA_ATTEMPTS})...")
+            logger.notice(f"Setting up Vespa (attempt {x+1}/{num_attempts})...")
             document_index.ensure_indices_exist(
                 index_embedding_dim=index_setting.model_dim,
                 secondary_index_embedding_dim=secondary_index_setting.model_dim
@@ -244,7 +245,7 @@ def setup_vespa(
             time.sleep(WAIT_SECONDS)
 
     logger.error(
-        f"Vespa setup did not succeed. Attempt limit reached. ({VESPA_ATTEMPTS})"
+        f"Vespa setup did not succeed. Attempt limit reached. ({num_attempts})"
     )
     return False
 

backend/danswer/tools/base_tool.py

@@ -7,7 +7,7 @@ from danswer.llm.utils import message_to_prompt_and_imgs
 from danswer.tools.tool import Tool
 
 if TYPE_CHECKING:
-    from danswer.llm.answering.prompts.build import AnswerPromptBuilder
+    from danswer.chat.prompt_builder.build import AnswerPromptBuilder
     from danswer.tools.tool_implementations.custom.custom_tool import (
         CustomToolCallSummary,
     )

backend/danswer/tools/message.py

@@ -25,9 +25,6 @@ class ToolCallSummary(BaseModel__v1):
     tool_call_request: AIMessage
     tool_call_result: ToolMessage
 
-    class Config:
-        arbitrary_types_allowed = True
-
 
 def tool_call_tokens(
     tool_call_summary: ToolCallSummary, llm_tokenizer: BaseTokenizer

backend/danswer/tools/tool.py

@@ -3,13 +3,13 @@ from collections.abc import Generator
 from typing import Any
 from typing import TYPE_CHECKING
 
-from danswer.llm.answering.models import PreviousMessage
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.utils.special_types import JSON_ro
 
 
 if TYPE_CHECKING:
-    from danswer.llm.answering.prompts.build import AnswerPromptBuilder
+    from danswer.chat.prompt_builder.build import AnswerPromptBuilder
     from danswer.tools.message import ToolCallSummary
     from danswer.tools.models import ToolResponse
 

backend/danswer/tools/tool_constructor.py

@@ -5,6 +5,10 @@ from pydantic import BaseModel
 from pydantic import Field
 from sqlalchemy.orm import Session
 
+from danswer.chat.models import AnswerStyleConfig
+from danswer.chat.models import CitationConfig
+from danswer.chat.models import DocumentPruningConfig
+from danswer.chat.models import PromptConfig
 from danswer.configs.app_configs import AZURE_DALLE_API_BASE
 from danswer.configs.app_configs import AZURE_DALLE_API_KEY
 from danswer.configs.app_configs import AZURE_DALLE_API_VERSION
@@ -19,10 +23,6 @@ from danswer.db.llm import fetch_existing_llm_providers
 from danswer.db.models import Persona
 from danswer.db.models import User
 from danswer.file_store.models import InMemoryChatFile
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import CitationConfig
-from danswer.llm.answering.models import DocumentPruningConfig
-from danswer.llm.answering.models import PromptConfig
 from danswer.llm.interfaces import LLM
 from danswer.llm.interfaces import LLMConfig
 from danswer.natural_language_processing.utils import get_tokenizer

backend/danswer/tools/tool_implementations/custom/custom_tool.py

@@ -15,14 +15,14 @@ from langchain_core.messages import SystemMessage
 from pydantic import BaseModel
 from requests import JSONDecodeError
 
+from danswer.chat.prompt_builder.build import AnswerPromptBuilder
 from danswer.configs.constants import FileOrigin
 from danswer.db.engine import get_session_with_default_tenant
 from danswer.file_store.file_store import get_default_file_store
 from danswer.file_store.models import ChatFileType
 from danswer.file_store.models import InMemoryChatFile
-from danswer.llm.answering.models import PreviousMessage
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.tools.base_tool import BaseTool
 from danswer.tools.message import ToolCallSummary
 from danswer.tools.models import CHAT_SESSION_ID_PLACEHOLDER

backend/danswer/tools/tool_implementations/images/image_generation_tool.py

@@ -4,14 +4,16 @@ from enum import Enum
 from typing import Any
 from typing import cast
 
+import requests
 from litellm import image_generation  # type: ignore
 from pydantic import BaseModel
 
 from danswer.chat.chat_utils import combine_message_chain
+from danswer.chat.prompt_builder.build import AnswerPromptBuilder
 from danswer.configs.model_configs import GEN_AI_HISTORY_CUTOFF
-from danswer.llm.answering.models import PreviousMessage
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
+from danswer.configs.tool_configs import IMAGE_GENERATION_OUTPUT_FORMAT
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.llm.utils import build_content_with_imgs
 from danswer.llm.utils import message_to_string
 from danswer.prompts.constants import GENERAL_SEP_PAT
@@ -56,9 +58,18 @@ Follow Up Input:
 """.strip()
 
 
+class ImageFormat(str, Enum):
+    URL = "url"
+    BASE64 = "b64_json"
+
+
+_DEFAULT_OUTPUT_FORMAT = ImageFormat(IMAGE_GENERATION_OUTPUT_FORMAT)
+
+
 class ImageGenerationResponse(BaseModel):
     revised_prompt: str
-    url: str
+    url: str | None
+    image_data: str | None
 
 
 class ImageShape(str, Enum):
@@ -80,6 +91,7 @@ class ImageGenerationTool(Tool):
         model: str = "dall-e-3",
         num_imgs: int = 2,
         additional_headers: dict[str, str] | None = None,
+        output_format: ImageFormat = _DEFAULT_OUTPUT_FORMAT,
     ) -> None:
         self.api_key = api_key
         self.api_base = api_base
@@ -89,6 +101,7 @@ class ImageGenerationTool(Tool):
         self.num_imgs = num_imgs
 
         self.additional_headers = additional_headers
+        self.output_format = output_format
 
     @property
     def name(self) -> str:
@@ -168,7 +181,7 @@ class ImageGenerationTool(Tool):
         )
 
         return build_content_with_imgs(
-            json.dumps(
+            message=json.dumps(
                 [
                     {
                         "revised_prompt": image_generation.revised_prompt,
@@ -177,13 +190,10 @@ class ImageGenerationTool(Tool):
                     for image_generation in image_generations
                 ]
             ),
-            # NOTE: we can't pass in the image URLs here, since OpenAI doesn't allow
-            # Tool messages to contain images
-            # img_urls=[image_generation.url for image_generation in image_generations],
         )
 
     def _generate_image(
-        self, prompt: str, shape: ImageShape
+        self, prompt: str, shape: ImageShape, format: ImageFormat
     ) -> ImageGenerationResponse:
         if shape == ImageShape.LANDSCAPE:
             size = "1792x1024"
@@ -197,20 +207,32 @@ class ImageGenerationTool(Tool):
                 prompt=prompt,
                 model=self.model,
                 api_key=self.api_key,
-                # need to pass in None rather than empty str
                 api_base=self.api_base or None,
                 api_version=self.api_version or None,
                 size=size,
                 n=1,
+                response_format=format,
                 extra_headers=build_llm_extra_headers(self.additional_headers),
             )
+
+            if format == ImageFormat.URL:
+                url = response.data[0]["url"]
+                image_data = None
+            else:
+                url = None
+                image_data = response.data[0]["b64_json"]
+
             return ImageGenerationResponse(
                 revised_prompt=response.data[0]["revised_prompt"],
-                url=response.data[0]["url"],
+                url=url,
+                image_data=image_data,
             )
 
+        except requests.RequestException as e:
+            logger.error(f"Error fetching or converting image: {e}")
+            raise ValueError("Failed to fetch or convert the generated image")
         except Exception as e:
-            logger.debug(f"Error occured during image generation: {e}")
+            logger.debug(f"Error occurred during image generation: {e}")
 
             error_message = str(e)
             if "OpenAIException" in str(type(e)):
@@ -235,9 +257,8 @@ class ImageGenerationTool(Tool):
     def run(self, **kwargs: str) -> Generator[ToolResponse, None, None]:
         prompt = cast(str, kwargs["prompt"])
         shape = ImageShape(kwargs.get("shape", ImageShape.SQUARE))
+        format = self.output_format
 
-        # dalle3 only supports 1 image at a time, which is why we have to
-        # parallelize this via threading
         results = cast(
             list[ImageGenerationResponse],
             run_functions_tuples_in_parallel(
@@ -247,6 +268,7 @@ class ImageGenerationTool(Tool):
                         (
                             prompt,
                             shape,
+                            format,
                         ),
                     )
                     for _ in range(self.num_imgs)
@@ -288,11 +310,17 @@ class ImageGenerationTool(Tool):
         if img_generation_response is None:
             raise ValueError("No image generation response found")
 
-        img_urls = [img.url for img in img_generation_response]
+        img_urls = [img.url for img in img_generation_response if img.url is not None]
+        b64_imgs = [
+            img.image_data
+            for img in img_generation_response
+            if img.image_data is not None
+        ]
         prompt_builder.update_user_prompt(
             build_image_generation_user_prompt(
                 query=prompt_builder.get_user_message_content(),
                 img_urls=img_urls,
+                b64_imgs=b64_imgs,
             )
         )
 

backend/danswer/tools/tool_implementations/images/prompt.py

@@ -11,11 +11,14 @@ Can you please summarize them in a sentence or two? Do NOT include image urls or
 
 
 def build_image_generation_user_prompt(
-    query: str, img_urls: list[str] | None = None
+    query: str,
+    img_urls: list[str] | None = None,
+    b64_imgs: list[str] | None = None,
 ) -> HumanMessage:
     return HumanMessage(
         content=build_content_with_imgs(
             message=IMG_GENERATION_SUMMARY_PROMPT.format(query=query).strip(),
+            b64_imgs=b64_imgs,
             img_urls=img_urls,
         )
     )

backend/danswer/tools/tool_implementations/internet_search/internet_search_tool.py

@@ -7,15 +7,15 @@ from typing import cast
 import httpx
 
 from danswer.chat.chat_utils import combine_message_chain
+from danswer.chat.models import AnswerStyleConfig
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import PromptConfig
+from danswer.chat.prompt_builder.build import AnswerPromptBuilder
 from danswer.configs.constants import DocumentSource
 from danswer.configs.model_configs import GEN_AI_HISTORY_CUTOFF
 from danswer.context.search.models import SearchDoc
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import PreviousMessage
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.llm.utils import message_to_string
 from danswer.prompts.chat_prompts import INTERNET_SEARCH_QUERY_REPHRASE
 from danswer.prompts.constants import GENERAL_SEP_PAT

backend/danswer/tools/tool_implementations/search/search_tool.py

@@ -7,10 +7,19 @@ from pydantic import BaseModel
 from sqlalchemy.orm import Session
 
 from danswer.chat.chat_utils import llm_doc_from_inference_section
+from danswer.chat.llm_response_handler import LLMCall
+from danswer.chat.models import AnswerStyleConfig
+from danswer.chat.models import ContextualPruningConfig
 from danswer.chat.models import DanswerContext
 from danswer.chat.models import DanswerContexts
+from danswer.chat.models import DocumentPruningConfig
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import PromptConfig
 from danswer.chat.models import SectionRelevancePiece
+from danswer.chat.prompt_builder.build import AnswerPromptBuilder
+from danswer.chat.prompt_builder.citations_prompt import compute_max_llm_input_tokens
+from danswer.chat.prune_and_merge import prune_and_merge_sections
+from danswer.chat.prune_and_merge import prune_sections
 from danswer.configs.chat_configs import CONTEXT_CHUNKS_ABOVE
 from danswer.configs.chat_configs import CONTEXT_CHUNKS_BELOW
 from danswer.configs.model_configs import GEN_AI_MODEL_FALLBACK_MAX_TOKENS
@@ -25,17 +34,8 @@ from danswer.context.search.models import SearchRequest
 from danswer.context.search.pipeline import SearchPipeline
 from danswer.db.models import Persona
 from danswer.db.models import User
-from danswer.llm.answering.llm_response_handler import LLMCall
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import ContextualPruningConfig
-from danswer.llm.answering.models import DocumentPruningConfig
-from danswer.llm.answering.models import PreviousMessage
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
-from danswer.llm.answering.prompts.citations_prompt import compute_max_llm_input_tokens
-from danswer.llm.answering.prune_and_merge import prune_and_merge_sections
-from danswer.llm.answering.prune_and_merge import prune_sections
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.secondary_llm_flows.choose_search import check_if_need_search
 from danswer.secondary_llm_flows.query_expansion import history_based_query_rephrase
 from danswer.tools.message import ToolCallSummary

backend/danswer/tools/tool_implementations/search_like_tool_utils.py

@@ -2,15 +2,15 @@ from typing import cast
 
 from langchain_core.messages import HumanMessage
 
+from danswer.chat.models import AnswerStyleConfig
 from danswer.chat.models import LlmDoc
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
-from danswer.llm.answering.prompts.citations_prompt import (
+from danswer.chat.models import PromptConfig
+from danswer.chat.prompt_builder.build import AnswerPromptBuilder
+from danswer.chat.prompt_builder.citations_prompt import (
     build_citations_system_message,
 )
-from danswer.llm.answering.prompts.citations_prompt import build_citations_user_message
-from danswer.llm.answering.prompts.quotes_prompt import build_quotes_user_message
+from danswer.chat.prompt_builder.citations_prompt import build_citations_user_message
+from danswer.chat.prompt_builder.quotes_prompt import build_quotes_user_message
 from danswer.tools.message import ToolCallSummary
 from danswer.tools.models import ToolResponse
 

backend/danswer/tools/tool_runner.py

@@ -2,8 +2,8 @@ from collections.abc import Callable
 from collections.abc import Generator
 from typing import Any
 
-from danswer.llm.answering.models import PreviousMessage
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.tools.models import ToolCallFinalResult
 from danswer.tools.models import ToolCallKickoff
 from danswer.tools.models import ToolResponse

backend/danswer/tools/tool_selection.py

@@ -3,8 +3,8 @@ from typing import Any
 
 from danswer.chat.chat_utils import combine_message_chain
 from danswer.configs.model_configs import GEN_AI_HISTORY_CUTOFF
-from danswer.llm.answering.models import PreviousMessage
 from danswer.llm.interfaces import LLM
+from danswer.llm.models import PreviousMessage
 from danswer.llm.utils import message_to_string
 from danswer.prompts.constants import GENERAL_SEP_PAT
 from danswer.tools.tool import Tool

backend/danswer/utils/b64.py

@@ -0,0 +1,25 @@
+import base64
+
+
+def get_image_type_from_bytes(raw_b64_bytes: bytes) -> str:
+    magic_number = raw_b64_bytes[:4]
+
+    if magic_number.startswith(b"\x89PNG"):
+        mime_type = "image/png"
+    elif magic_number.startswith(b"\xFF\xD8"):
+        mime_type = "image/jpeg"
+    elif magic_number.startswith(b"GIF8"):
+        mime_type = "image/gif"
+    elif magic_number.startswith(b"RIFF") and raw_b64_bytes[8:12] == b"WEBP":
+        mime_type = "image/webp"
+    else:
+        raise ValueError(
+            "Unsupported image format - only PNG, JPEG, " "GIF, and WEBP are supported."
+        )
+
+    return mime_type
+
+
+def get_image_type(raw_b64_string: str) -> str:
+    binary_data = base64.b64decode(raw_b64_string)
+    return get_image_type_from_bytes(binary_data)

backend/danswer/utils/timing.py

@@ -33,12 +33,12 @@ def log_function_time(
             elapsed_time_str = f"{elapsed_time:.3f}"
             log_name = func_name or func.__name__
             args_str = f" args={args} kwargs={kwargs}" if include_args else ""
-            f"{log_name}{args_str} took {elapsed_time_str} seconds"
-            # if debug_only:
-            #     logger.debug(final_log)
-            # else:
-            #     # These are generally more important logs so the level is a bit higher
-            #     logger.notice(final_log)
+            final_log = f"{log_name}{args_str} took {elapsed_time_str} seconds"
+            if debug_only:
+                logger.debug(final_log)
+            else:
+                # These are generally more important logs so the level is a bit higher
+                logger.notice(final_log)
 
             if not print_only:
                 optional_telemetry(

backend/ee/danswer/configs/app_configs.py

@@ -11,6 +11,14 @@ SAML_CONF_DIR = os.environ.get("SAML_CONF_DIR") or "/app/ee/danswer/configs/saml
 #####
 # Auto Permission Sync
 #####
+# In seconds, default is 5 minutes
+CONFLUENCE_PERMISSION_GROUP_SYNC_FREQUENCY = int(
+    os.environ.get("CONFLUENCE_PERMISSION_GROUP_SYNC_FREQUENCY") or 5 * 60
+)
+# In seconds, default is 5 minutes
+CONFLUENCE_PERMISSION_DOC_SYNC_FREQUENCY = int(
+    os.environ.get("CONFLUENCE_PERMISSION_DOC_SYNC_FREQUENCY") or 5 * 60
+)
 NUM_PERMISSION_WORKERS = int(os.environ.get("NUM_PERMISSION_WORKERS") or 2)
 
 

backend/ee/danswer/db/analytics.py

@@ -170,3 +170,67 @@ def fetch_danswerbot_analytics(
     )
 
     return results
+
+
+def fetch_persona_message_analytics(
+    db_session: Session,
+    persona_id: int,
+    start: datetime.datetime,
+    end: datetime.datetime,
+) -> list[tuple[int, datetime.date]]:
+    """Gets the daily message counts for a specific persona within the given time range."""
+    query = (
+        select(
+            func.count(ChatMessage.id),
+            cast(ChatMessage.time_sent, Date),
+        )
+        .join(
+            ChatSession,
+            ChatMessage.chat_session_id == ChatSession.id,
+        )
+        .where(
+            or_(
+                ChatMessage.alternate_assistant_id == persona_id,
+                ChatSession.persona_id == persona_id,
+            ),
+            ChatMessage.time_sent >= start,
+            ChatMessage.time_sent <= end,
+            ChatMessage.message_type == MessageType.ASSISTANT,
+        )
+        .group_by(cast(ChatMessage.time_sent, Date))
+        .order_by(cast(ChatMessage.time_sent, Date))
+    )
+
+    return [tuple(row) for row in db_session.execute(query).all()]
+
+
+def fetch_persona_unique_users(
+    db_session: Session,
+    persona_id: int,
+    start: datetime.datetime,
+    end: datetime.datetime,
+) -> list[tuple[int, datetime.date]]:
+    """Gets the daily unique user counts for a specific persona within the given time range."""
+    query = (
+        select(
+            func.count(func.distinct(ChatSession.user_id)),
+            cast(ChatMessage.time_sent, Date),
+        )
+        .join(
+            ChatSession,
+            ChatMessage.chat_session_id == ChatSession.id,
+        )
+        .where(
+            or_(
+                ChatMessage.alternate_assistant_id == persona_id,
+                ChatSession.persona_id == persona_id,
+            ),
+            ChatMessage.time_sent >= start,
+            ChatMessage.time_sent <= end,
+            ChatMessage.message_type == MessageType.ASSISTANT,
+        )
+        .group_by(cast(ChatMessage.time_sent, Date))
+        .order_by(cast(ChatMessage.time_sent, Date))
+    )
+
+    return [tuple(row) for row in db_session.execute(query).all()]

backend/ee/danswer/db/external_perm.py

@@ -10,6 +10,9 @@ from danswer.access.utils import prefix_group_w_source
 from danswer.configs.constants import DocumentSource
 from danswer.db.models import User__ExternalUserGroupId
 from danswer.db.users import batch_add_ext_perm_user_if_not_exists
+from danswer.utils.logger import setup_logger
+
+logger = setup_logger()
 
 
 class ExternalUserGroup(BaseModel):
@@ -73,7 +76,13 @@ def replace_user__ext_group_for_cc_pair(
     new_external_permissions = []
     for external_group in group_defs:
         for user_email in external_group.user_emails:
-            user_id = email_id_map[user_email]
+            user_id = email_id_map.get(user_email)
+            if user_id is None:
+                logger.warning(
+                    f"User in group {external_group.id}"
+                    f" with email {user_email} not found"
+                )
+                continue
             new_external_permissions.append(
                 User__ExternalUserGroupId(
                     user_id=user_id,

backend/ee/danswer/external_permissions/confluence/doc_sync.py

@@ -195,6 +195,7 @@ def _fetch_all_page_restrictions_for_space(
     confluence_client: OnyxConfluence,
     slim_docs: list[SlimDocument],
     space_permissions_by_space_key: dict[str, ExternalAccess],
+    is_cloud: bool,
 ) -> list[DocExternalAccess]:
     """
     For all pages, if a page has restrictions, then use those restrictions.
@@ -222,27 +223,50 @@ def _fetch_all_page_restrictions_for_space(
             continue
 
         space_key = slim_doc.perm_sync_data.get("space_key")
-        if space_permissions := space_permissions_by_space_key.get(space_key):
-            # If there are no restrictions, then use the space's restrictions
-            document_restrictions.append(
-                DocExternalAccess(
-                    doc_id=slim_doc.id,
-                    external_access=space_permissions,
-                )
+        if not (space_permissions := space_permissions_by_space_key.get(space_key)):
+            logger.debug(
+                f"Individually fetching space permissions for space {space_key}"
             )
-            if (
-                not space_permissions.is_public
-                and not space_permissions.external_user_emails
-                and not space_permissions.external_user_group_ids
-            ):
+            try:
+                # If the space permissions are not in the cache, then fetch them
+                if is_cloud:
+                    retrieved_space_permissions = _get_cloud_space_permissions(
+                        confluence_client=confluence_client, space_key=space_key
+                    )
+                else:
+                    retrieved_space_permissions = _get_server_space_permissions(
+                        confluence_client=confluence_client, space_key=space_key
+                    )
+                space_permissions_by_space_key[space_key] = retrieved_space_permissions
+                space_permissions = retrieved_space_permissions
+            except Exception as e:
                 logger.warning(
-                    f"Permissions are empty for document: {slim_doc.id}\n"
-                    "This means space permissions are may be wrong for"
-                    f" Space key: {space_key}"
+                    f"Error fetching space permissions for space {space_key}: {e}"
                 )
+
+        if not space_permissions:
+            logger.warning(
+                f"No permissions found for document {slim_doc.id} in space {space_key}"
+            )
             continue
 
-        logger.warning(f"No permissions found for document {slim_doc.id}")
+        # If there are no restrictions, then use the space's restrictions
+        document_restrictions.append(
+            DocExternalAccess(
+                doc_id=slim_doc.id,
+                external_access=space_permissions,
+            )
+        )
+        if (
+            not space_permissions.is_public
+            and not space_permissions.external_user_emails
+            and not space_permissions.external_user_group_ids
+        ):
+            logger.warning(
+                f"Permissions are empty for document: {slim_doc.id}\n"
+                "This means space permissions are may be wrong for"
+                f" Space key: {space_key}"
+            )
 
     logger.debug("Finished fetching all page restrictions for space")
     return document_restrictions
@@ -281,4 +305,5 @@ def confluence_doc_sync(
         confluence_client=confluence_connector.confluence_client,
         slim_docs=slim_docs,
         space_permissions_by_space_key=space_permissions_by_space_key,
+        is_cloud=is_cloud,
     )

backend/ee/danswer/external_permissions/sync_params.py

@@ -3,6 +3,8 @@ from collections.abc import Callable
 from danswer.access.models import DocExternalAccess
 from danswer.configs.constants import DocumentSource
 from danswer.db.models import ConnectorCredentialPair
+from ee.danswer.configs.app_configs import CONFLUENCE_PERMISSION_DOC_SYNC_FREQUENCY
+from ee.danswer.configs.app_configs import CONFLUENCE_PERMISSION_GROUP_SYNC_FREQUENCY
 from ee.danswer.db.external_perm import ExternalUserGroup
 from ee.danswer.external_permissions.confluence.doc_sync import confluence_doc_sync
 from ee.danswer.external_permissions.confluence.group_sync import confluence_group_sync
@@ -56,7 +58,7 @@ GROUP_PERMISSIONS_IS_CC_PAIR_AGNOSTIC: set[DocumentSource] = {
 # If nothing is specified here, we run the doc_sync every time the celery beat runs
 DOC_PERMISSION_SYNC_PERIODS: dict[DocumentSource, int] = {
     # Polling is not supported so we fetch all doc permissions every 5 minutes
-    DocumentSource.CONFLUENCE: 5 * 60,
+    DocumentSource.CONFLUENCE: CONFLUENCE_PERMISSION_DOC_SYNC_FREQUENCY,
     DocumentSource.SLACK: 5 * 60,
 }
 
@@ -64,7 +66,7 @@ DOC_PERMISSION_SYNC_PERIODS: dict[DocumentSource, int] = {
 EXTERNAL_GROUP_SYNC_PERIODS: dict[DocumentSource, int] = {
     # Polling is not supported so we fetch all group permissions every 30 minutes
     DocumentSource.GOOGLE_DRIVE: 5 * 60,
-    DocumentSource.CONFLUENCE: 30 * 60,
+    DocumentSource.CONFLUENCE: CONFLUENCE_PERMISSION_GROUP_SYNC_FREQUENCY,
 }
 
 

backend/ee/danswer/server/analytics/api.py

@@ -11,11 +11,16 @@ from danswer.db.engine import get_session
 from danswer.db.models import User
 from ee.danswer.db.analytics import fetch_danswerbot_analytics
 from ee.danswer.db.analytics import fetch_per_user_query_analytics
+from ee.danswer.db.analytics import fetch_persona_message_analytics
+from ee.danswer.db.analytics import fetch_persona_unique_users
 from ee.danswer.db.analytics import fetch_query_analytics
 
 router = APIRouter(prefix="/analytics")
 
 
+_DEFAULT_LOOKBACK_DAYS = 30
+
+
 class QueryAnalyticsResponse(BaseModel):
     total_queries: int
     total_likes: int
@@ -33,7 +38,7 @@ def get_query_analytics(
     daily_query_usage_info = fetch_query_analytics(
         start=start
         or (
-            datetime.datetime.utcnow() - datetime.timedelta(days=30)
+            datetime.datetime.utcnow() - datetime.timedelta(days=_DEFAULT_LOOKBACK_DAYS)
         ),  # default is 30d lookback
         end=end or datetime.datetime.utcnow(),
         db_session=db_session,
@@ -64,7 +69,7 @@ def get_user_analytics(
     daily_query_usage_info_per_user = fetch_per_user_query_analytics(
         start=start
         or (
-            datetime.datetime.utcnow() - datetime.timedelta(days=30)
+            datetime.datetime.utcnow() - datetime.timedelta(days=_DEFAULT_LOOKBACK_DAYS)
         ),  # default is 30d lookback
         end=end or datetime.datetime.utcnow(),
         db_session=db_session,
@@ -98,7 +103,7 @@ def get_danswerbot_analytics(
     daily_danswerbot_info = fetch_danswerbot_analytics(
         start=start
         or (
-            datetime.datetime.utcnow() - datetime.timedelta(days=30)
+            datetime.datetime.utcnow() - datetime.timedelta(days=_DEFAULT_LOOKBACK_DAYS)
         ),  # default is 30d lookback
         end=end or datetime.datetime.utcnow(),
         db_session=db_session,
@@ -115,3 +120,74 @@ def get_danswerbot_analytics(
     ]
 
     return resolution_results
+
+
+class PersonaMessageAnalyticsResponse(BaseModel):
+    total_messages: int
+    date: datetime.date
+    persona_id: int
+
+
+@router.get("/admin/persona/messages")
+def get_persona_messages(
+    persona_id: int,
+    start: datetime.datetime | None = None,
+    end: datetime.datetime | None = None,
+    _: User | None = Depends(current_admin_user),
+    db_session: Session = Depends(get_session),
+) -> list[PersonaMessageAnalyticsResponse]:
+    """Fetch daily message counts for a single persona within the given time range."""
+    start = start or (
+        datetime.datetime.utcnow() - datetime.timedelta(days=_DEFAULT_LOOKBACK_DAYS)
+    )
+    end = end or datetime.datetime.utcnow()
+
+    persona_message_counts = []
+    for count, date in fetch_persona_message_analytics(
+        db_session=db_session,
+        persona_id=persona_id,
+        start=start,
+        end=end,
+    ):
+        persona_message_counts.append(
+            PersonaMessageAnalyticsResponse(
+                total_messages=count,
+                date=date,
+                persona_id=persona_id,
+            )
+        )
+
+    return persona_message_counts
+
+
+class PersonaUniqueUsersResponse(BaseModel):
+    unique_users: int
+    date: datetime.date
+    persona_id: int
+
+
+@router.get("/admin/persona/unique-users")
+def get_persona_unique_users(
+    persona_id: int,
+    start: datetime.datetime,
+    end: datetime.datetime,
+    _: User | None = Depends(current_admin_user),
+    db_session: Session = Depends(get_session),
+) -> list[PersonaUniqueUsersResponse]:
+    """Get unique users per day for a single persona."""
+    unique_user_counts = []
+    daily_counts = fetch_persona_unique_users(
+        db_session=db_session,
+        persona_id=persona_id,
+        start=start,
+        end=end,
+    )
+    for count, date in daily_counts:
+        unique_user_counts.append(
+            PersonaUniqueUsersResponse(
+                unique_users=count,
+                date=date,
+                persona_id=persona_id,
+            )
+        )
+    return unique_user_counts

backend/requirements/default.txt

@@ -26,14 +26,9 @@ huggingface-hub==0.20.1
 jira==3.5.1
 jsonref==1.1.0
 trafilatura==1.12.2
-langchain==0.3.7
-langchain-core==0.3.24
-langchain-openai==0.2.9
-langchain-text-splitters==0.3.2
-langchainhub==0.1.21
-langgraph==0.2.59
-langgraph-checkpoint==2.0.5
-langgraph-sdk==0.1.44
+langchain==0.1.17
+langchain-core==0.1.50
+langchain-text-splitters==0.0.1
 litellm==1.53.1
 lxml==5.3.0
 lxml_html_clean==0.2.2

backend/shared_configs/configs.py

@@ -61,7 +61,7 @@ LOG_FILE_NAME = os.environ.get("LOG_FILE_NAME") or "danswer"
 # Enable generating persistent log files for local dev environments
 DEV_LOGGING_ENABLED = os.environ.get("DEV_LOGGING_ENABLED", "").lower() == "true"
 # notset, debug, info, notice, warning, error, or critical
-LOG_LEVEL = os.environ.get("LOG_LEVEL", "notice")
+LOG_LEVEL = os.environ.get("LOG_LEVEL", "info")
 
 # Timeout for API-based embedding models
 # NOTE: does not apply for Google VertexAI, since the python client doesn't

backend/tests/daily/connectors/confluence/test_confluence_permissions_basic.py

@@ -0,0 +1,39 @@
+import os
+
+import pytest
+
+from danswer.connectors.confluence.connector import ConfluenceConnector
+
+
+@pytest.fixture
+def confluence_connector() -> ConfluenceConnector:
+    connector = ConfluenceConnector(
+        wiki_base="https://danswerai.atlassian.net",
+        is_cloud=True,
+    )
+    connector.load_credentials(
+        {
+            "confluence_access_token": os.environ["CONFLUENCE_ACCESS_TOKEN"],
+            "confluence_username": os.environ["CONFLUENCE_USER_NAME"],
+        }
+    )
+    return connector
+
+
+# This should never fail because even if the docs in the cloud change,
+# the full doc ids retrieved should always be a subset of the slim doc ids
+def test_confluence_connector_permissions(
+    confluence_connector: ConfluenceConnector,
+) -> None:
+    # Get all doc IDs from the full connector
+    all_full_doc_ids = set()
+    for doc_batch in confluence_connector.load_from_state():
+        all_full_doc_ids.update([doc.id for doc in doc_batch])
+
+    # Get all doc IDs from the slim connector
+    all_slim_doc_ids = set()
+    for slim_doc_batch in confluence_connector.retrieve_all_slim_documents():
+        all_slim_doc_ids.update([doc.id for doc in slim_doc_batch])
+
+    # The set of full doc IDs should be always be a subset of the slim doc IDs
+    assert all_full_doc_ids.issubset(all_slim_doc_ids)

backend/tests/unit/danswer/chat/conftest.py

@@ -5,12 +5,12 @@ from unittest.mock import MagicMock
 import pytest
 from langchain_core.messages import SystemMessage
 
+from danswer.chat.models import AnswerStyleConfig
+from danswer.chat.models import CitationConfig
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import PromptConfig
+from danswer.chat.prompt_builder.build import AnswerPromptBuilder
 from danswer.configs.constants import DocumentSource
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import CitationConfig
-from danswer.llm.answering.models import PromptConfig
-from danswer.llm.answering.prompts.build import AnswerPromptBuilder
 from danswer.llm.interfaces import LLMConfig
 from danswer.tools.models import ToolResponse
 from danswer.tools.tool_implementations.search.search_tool import SearchTool

backend/tests/unit/danswer/chat/stream_processing/test_citation_processing.py

@@ -5,11 +5,9 @@ import pytest
 from danswer.chat.models import CitationInfo
 from danswer.chat.models import DanswerAnswerPiece
 from danswer.chat.models import LlmDoc
+from danswer.chat.stream_processing.citation_processing import CitationProcessor
+from danswer.chat.stream_processing.utils import DocumentIdOrderMapping
 from danswer.configs.constants import DocumentSource
-from danswer.llm.answering.stream_processing.citation_processing import (
-    CitationProcessor,
-)
-from danswer.llm.answering.stream_processing.utils import DocumentIdOrderMapping
 
 
 """

backend/tests/unit/danswer/chat/stream_processing/test_quotes_processing.py

@@ -2,14 +2,10 @@ import textwrap
 
 import pytest
 
+from danswer.chat.stream_processing.quotes_processing import match_quotes_to_docs
+from danswer.chat.stream_processing.quotes_processing import separate_answer_quotes
 from danswer.configs.constants import DocumentSource
 from danswer.context.search.models import InferenceChunk
-from danswer.llm.answering.stream_processing.quotes_processing import (
-    match_quotes_to_docs,
-)
-from danswer.llm.answering.stream_processing.quotes_processing import (
-    separate_answer_quotes,
-)
 
 
 def test_passed_in_quotes() -> None:

backend/tests/unit/danswer/chat/test_answer.py

@@ -11,21 +11,21 @@ from langchain_core.messages import SystemMessage
 from langchain_core.messages import ToolCall
 from langchain_core.messages import ToolCallChunk
 
+from danswer.chat.answer import Answer
+from danswer.chat.models import AnswerStyleConfig
 from danswer.chat.models import CitationInfo
 from danswer.chat.models import DanswerAnswerPiece
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import PromptConfig
 from danswer.chat.models import StreamStopInfo
 from danswer.chat.models import StreamStopReason
-from danswer.llm.answering.answer import Answer
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import PromptConfig
 from danswer.llm.interfaces import LLM
 from danswer.tools.force import ForceUseTool
 from danswer.tools.models import ToolCallFinalResult
 from danswer.tools.models import ToolCallKickoff
 from danswer.tools.models import ToolResponse
-from tests.unit.danswer.llm.answering.conftest import DEFAULT_SEARCH_ARGS
-from tests.unit.danswer.llm.answering.conftest import QUERY
+from tests.unit.danswer.chat.conftest import DEFAULT_SEARCH_ARGS
+from tests.unit.danswer.chat.conftest import QUERY
 
 
 @pytest.fixture

backend/tests/unit/danswer/chat/test_prune_and_merge.py

@@ -1,9 +1,9 @@
 import pytest
 
+from danswer.chat.prune_and_merge import _merge_sections
 from danswer.configs.constants import DocumentSource
 from danswer.context.search.models import InferenceChunk
 from danswer.context.search.models import InferenceSection
-from danswer.llm.answering.prune_and_merge import _merge_sections
 
 
 # This large test accounts for all of the following:

backend/tests/unit/danswer/chat/test_skip_gen_ai.py

@@ -5,10 +5,10 @@ from unittest.mock import Mock
 import pytest
 from pytest_mock import MockerFixture
 
-from danswer.llm.answering.answer import Answer
-from danswer.llm.answering.answer import AnswerStream
-from danswer.llm.answering.models import AnswerStyleConfig
-from danswer.llm.answering.models import PromptConfig
+from danswer.chat.answer import Answer
+from danswer.chat.answer import AnswerStream
+from danswer.chat.models import AnswerStyleConfig
+from danswer.chat.models import PromptConfig
 from danswer.tools.force import ForceUseTool
 from danswer.tools.tool_implementations.search.search_tool import SearchTool
 from tests.regression.answer_quality.run_qa import _process_and_write_query_results