.

Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>

.cursor/skills/onyx-cli/SKILL.md

@@ -106,13 +106,34 @@ onyx-cli ask --json "What authentication methods do we support?"
 
 Outputs JSON-encoded parsed stream events (one object per line). Key event objects include message deltas, stop, errors, search-start, and citation payloads.
 
+Each line is a JSON object with this envelope:
+
+```json
+{"type": "<event_type>", "event": { ... }}
+```
+
 | Event Type | Description |
 |------------|-------------|
 | `message_delta` | Content token — concatenate all `content` fields for the full answer |
 | `stop` | Stream complete |
 | `error` | Error with `error` message field |
 | `search_tool_start` | Onyx started searching documents |
-| `citation_info` | Source citation with `citation_number` and `document_id` |
+| `citation_info` | Source citation — see shape below |
+
+`citation_info` event shape:
+
+```json
+{
+  "type": "citation_info",
+  "event": {
+    "citation_number": 1,
+    "document_id": "abc123def456",
+    "placement": {"turn_index": 0, "tab_index": 0, "sub_turn_index": null}
+  }
+}
+```
+
+`placement` is metadata about where in the conversation the citation appeared and can be ignored for most use cases.
 
 ### Specify an agent
 
@@ -129,6 +150,10 @@ Uses a specific Onyx agent/persona instead of the default.
 | `--agent-id` | int | Agent ID to use (overrides default) |
 | `--json` | bool | Output raw NDJSON events instead of plain text |
 
+## Statelessness
+
+Each `onyx-cli ask` call creates an independent chat session. There is no built-in way to chain context across multiple `ask` invocations — every call starts fresh. If you need multi-turn conversation with memory, use the interactive TUI (`onyx-cli` or `onyx-cli chat`) instead.
+
 ## When to Use
 
 Use `onyx-cli ask` when:

.github/workflows/pr-desktop-build.yml

@@ -57,7 +57,7 @@ jobs:
           cache-dependency-path: ./desktop/package-lock.json
 
       - name: Setup Rust
-        uses: dtolnay/rust-toolchain@4be9e76fd7c4901c61fb841f559994984270fce7
+        uses: dtolnay/rust-toolchain@efa25f7f19611383d5b0ccf2d1c8914531636bf9
         with:
           toolchain: stable
           targets: ${{ matrix.target }}

.github/workflows/release-cli.yml

@@ -26,7 +26,6 @@ jobs:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # ratchet:actions/checkout@v6
         with:
           persist-credentials: false
-          fetch-depth: 0
       - uses: astral-sh/setup-uv@61cb8a9741eeb8a550a1b8544337180c0fc8476b # ratchet:astral-sh/setup-uv@v7
         with:
           enable-cache: false

.github/workflows/release-devtools.yml

@@ -22,12 +22,10 @@ jobs:
           - { goos: "windows", goarch: "arm64" }
           - { goos: "darwin", goarch: "amd64" }
           - { goos: "darwin", goarch: "arm64" }
-          - { goos: "", goarch: "" }
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # ratchet:actions/checkout@v6
         with:
           persist-credentials: false
-          fetch-depth: 0
       - uses: astral-sh/setup-uv@61cb8a9741eeb8a550a1b8544337180c0fc8476b # ratchet:astral-sh/setup-uv@v7
         with:
           enable-cache: false

AGENTS.md

@@ -544,6 +544,8 @@ To run them:
 npx playwright test <TEST_NAME>
 ```
 
+For shared fixtures, best practices, and detailed guidance, see `backend/tests/README.md`.
+
 ## Logs
 
 When (1) writing integration tests or (2) doing live tests (e.g. curl / playwright) you can get access

backend/ee/onyx/utils/encryption.py

@@ -14,67 +14,82 @@ from onyx.utils.variable_functionality import fetch_versioned_implementation
 logger = setup_logger()
 
 
-@lru_cache(maxsize=1)
+@lru_cache(maxsize=2)
 def _get_trimmed_key(key: str) -> bytes:
     encoded_key = key.encode()
     key_length = len(encoded_key)
     if key_length < 16:
         raise RuntimeError("Invalid ENCRYPTION_KEY_SECRET - too short")
-    elif key_length > 32:
-        key = key[:32]
-    elif key_length not in (16, 24, 32):
-        valid_lengths = [16, 24, 32]
-        key = key[: min(valid_lengths, key=lambda x: abs(x - key_length))]
 
-    return encoded_key
+    # Trim to the largest valid AES key size that fits
+    valid_lengths = [32, 24, 16]
+    for size in valid_lengths:
+        if key_length >= size:
+            return encoded_key[:size]
 
 
-def _encrypt_string(input_str: str) -> bytes:
-    if not ENCRYPTION_KEY_SECRET:
+def _encrypt_string(input_str: str, key: str | None = None) -> bytes:
+    effective_key = key if key is not None else ENCRYPTION_KEY_SECRET
+    if not effective_key:
         return input_str.encode()
 
-    key = _get_trimmed_key(ENCRYPTION_KEY_SECRET)
+    trimmed = _get_trimmed_key(effective_key)
     iv = urandom(16)
     padder = padding.PKCS7(algorithms.AES.block_size).padder()
     padded_data = padder.update(input_str.encode()) + padder.finalize()
 
-    cipher = Cipher(algorithms.AES(key), modes.CBC(iv), backend=default_backend())
+    cipher = Cipher(algorithms.AES(trimmed), modes.CBC(iv), backend=default_backend())
     encryptor = cipher.encryptor()
     encrypted_data = encryptor.update(padded_data) + encryptor.finalize()
 
     return iv + encrypted_data
 
 
-def _decrypt_bytes(input_bytes: bytes) -> str:
-    if not ENCRYPTION_KEY_SECRET:
+def _decrypt_bytes(input_bytes: bytes, key: str | None = None) -> str:
+    effective_key = key if key is not None else ENCRYPTION_KEY_SECRET
+    if not effective_key:
         return input_bytes.decode()
 
-    key = _get_trimmed_key(ENCRYPTION_KEY_SECRET)
-    iv = input_bytes[:16]
-    encrypted_data = input_bytes[16:]
+    trimmed = _get_trimmed_key(effective_key)
+    try:
+        iv = input_bytes[:16]
+        encrypted_data = input_bytes[16:]
 
-    cipher = Cipher(algorithms.AES(key), modes.CBC(iv), backend=default_backend())
-    decryptor = cipher.decryptor()
-    decrypted_padded_data = decryptor.update(encrypted_data) + decryptor.finalize()
+        cipher = Cipher(
+            algorithms.AES(trimmed), modes.CBC(iv), backend=default_backend()
+        )
+        decryptor = cipher.decryptor()
+        decrypted_padded_data = decryptor.update(encrypted_data) + decryptor.finalize()
 
-    unpadder = padding.PKCS7(algorithms.AES.block_size).unpadder()
-    decrypted_data = unpadder.update(decrypted_padded_data) + unpadder.finalize()
+        unpadder = padding.PKCS7(algorithms.AES.block_size).unpadder()
+        decrypted_data = unpadder.update(decrypted_padded_data) + unpadder.finalize()
 
-    return decrypted_data.decode()
+        return decrypted_data.decode()
+    except (ValueError, UnicodeDecodeError):
+        if key is not None:
+            # Explicit key was provided — don't fall back silently
+            raise
+        # Read path: fall back to raw decode for key rotation compatibility.
+        # Run the re-encrypt-secrets script to rotate to the current key.
+        logger.warning(
+            "AES decryption failed — falling back to raw decode. "
+            "Run the re-encrypt secrets script to rotate to the current key."
+        )
+        return input_bytes.decode()
 
 
-def encrypt_string_to_bytes(input_str: str) -> bytes:
+def encrypt_string_to_bytes(input_str: str, key: str | None = None) -> bytes:
     versioned_encryption_fn = fetch_versioned_implementation(
         "onyx.utils.encryption", "_encrypt_string"
     )
-    return versioned_encryption_fn(input_str)
+    return versioned_encryption_fn(input_str, key=key)
 
 
-def decrypt_bytes_to_string(input_bytes: bytes) -> str:
+def decrypt_bytes_to_string(input_bytes: bytes, key: str | None = None) -> str:
     versioned_decryption_fn = fetch_versioned_implementation(
         "onyx.utils.encryption", "_decrypt_bytes"
     )
-    return versioned_decryption_fn(input_bytes)
+    return versioned_decryption_fn(input_bytes, key=key)
 
 
 def test_encryption() -> None:

backend/onyx/db/rotate_encryption_key.py

@@ -0,0 +1,157 @@
+"""Rotate encryption key for all encrypted columns.
+
+Dynamically discovers all columns using EncryptedString / EncryptedJson,
+decrypts each value with the old key, and re-encrypts with the current
+ENCRYPTION_KEY_SECRET.
+
+The operation is idempotent: rows already encrypted with the current key
+are skipped. Commits are made in batches so a crash mid-rotation can be
+safely resumed by re-running.
+"""
+
+import json
+from typing import Any
+
+from sqlalchemy import LargeBinary
+from sqlalchemy import select
+from sqlalchemy import update
+from sqlalchemy.orm import Session
+
+from onyx.configs.app_configs import ENCRYPTION_KEY_SECRET
+from onyx.db.models import Base
+from onyx.db.models import EncryptedJson
+from onyx.db.models import EncryptedString
+from onyx.utils.encryption import decrypt_bytes_to_string
+from onyx.utils.logger import setup_logger
+
+logger = setup_logger()
+
+_BATCH_SIZE = 500
+
+
+def _can_decrypt_with_current_key(data: bytes) -> bool:
+    """Check if data is already encrypted with the current key.
+
+    Passes the key explicitly so the fallback-to-raw-decode path in
+    _decrypt_bytes is NOT triggered — a clean success/failure signal.
+    """
+    try:
+        decrypt_bytes_to_string(data, key=ENCRYPTION_KEY_SECRET)
+        return True
+    except Exception:
+        return False
+
+
+def _discover_encrypted_columns() -> list[tuple[type, str, list[str], bool]]:
+    """Walk all ORM models and find columns using EncryptedString/EncryptedJson.
+
+    Returns list of (ModelClass, column_attr_name, [pk_attr_names], is_json).
+    """
+    results: list[tuple[type, str, list[str], bool]] = []
+
+    for mapper in Base.registry.mappers:
+        model_cls = mapper.class_
+        pk_names = [col.key for col in mapper.primary_key]
+
+        for prop in mapper.column_attrs:
+            for col in prop.columns:
+                if isinstance(col.type, EncryptedJson):
+                    results.append((model_cls, prop.key, pk_names, True))
+                elif isinstance(col.type, EncryptedString):
+                    results.append((model_cls, prop.key, pk_names, False))
+
+    return results
+
+
+def rotate_encryption_key(
+    db_session: Session,
+    old_key: str | None,
+    dry_run: bool = False,
+) -> dict[str, int]:
+    """Decrypt all encrypted columns with old_key and re-encrypt with the current key.
+
+    Args:
+        db_session: Active database session.
+        old_key: The previous encryption key. Pass None or "" if values were
+                 not previously encrypted with a key.
+        dry_run: If True, count rows that need rotation without modifying data.
+
+    Returns:
+        Dict of "table.column" -> number of rows re-encrypted (or would be).
+
+    Commits every _BATCH_SIZE rows so that locks are held briefly and progress
+    is preserved on crash. Already-rotated rows are detected and skipped,
+    making the operation safe to re-run.
+    """
+    if not ENCRYPTION_KEY_SECRET:
+        raise RuntimeError(
+            "ENCRYPTION_KEY_SECRET is not set — cannot rotate. "
+            "Set the target encryption key in the environment before running."
+        )
+
+    encrypted_columns = _discover_encrypted_columns()
+    totals: dict[str, int] = {}
+
+    for model_cls, col_name, pk_names, is_json in encrypted_columns:
+        table_name: str = model_cls.__tablename__  # type: ignore[attr-defined]
+        col_attr = getattr(model_cls, col_name)
+        pk_attrs = [getattr(model_cls, pk) for pk in pk_names]
+
+        # Read raw bytes directly, bypassing the TypeDecorator
+        raw_col = col_attr.property.columns[0]
+
+        stmt = select(*pk_attrs, raw_col.cast(LargeBinary)).where(col_attr.is_not(None))
+        rows = db_session.execute(stmt).all()
+
+        reencrypted = 0
+        batch_pending = 0
+        for row in rows:
+            raw_bytes: bytes | None = row[-1]
+            if raw_bytes is None:
+                continue
+
+            if _can_decrypt_with_current_key(raw_bytes):
+                continue
+
+            try:
+                if not old_key:
+                    decrypted_str = raw_bytes.decode("utf-8")
+                else:
+                    decrypted_str = decrypt_bytes_to_string(raw_bytes, key=old_key)
+            except (ValueError, UnicodeDecodeError) as e:
+                pk_vals = [row[i] for i in range(len(pk_names))]
+                logger.warning(
+                    f"Could not decrypt {table_name}.{col_name} "
+                    f"row {pk_vals} — skipping: {e}"
+                )
+                continue
+
+            if not dry_run:
+                # For EncryptedJson, parse back to dict so the TypeDecorator
+                # can json.dumps() it cleanly (avoids double-encoding).
+                value: Any = json.loads(decrypted_str) if is_json else decrypted_str
+
+                pk_filters = [pk_attr == row[i] for i, pk_attr in enumerate(pk_attrs)]
+                update_stmt = (
+                    update(model_cls).where(*pk_filters).values({col_name: value})
+                )
+                db_session.execute(update_stmt)
+                batch_pending += 1
+
+                if batch_pending >= _BATCH_SIZE:
+                    db_session.commit()
+                    batch_pending = 0
+            reencrypted += 1
+
+        # Flush remaining rows in this column
+        if batch_pending > 0:
+            db_session.commit()
+
+        if reencrypted > 0:
+            totals[f"{table_name}.{col_name}"] = reencrypted
+            logger.info(
+                f"{'[DRY RUN] Would re-encrypt' if dry_run else 'Re-encrypted'} "
+                f"{reencrypted} value(s) in {table_name}.{col_name}"
+            )
+
+    return totals

backend/onyx/document_index/FILTER_SEMANTICS.md

@@ -0,0 +1,103 @@
+# Vector DB Filter Semantics
+
+How `IndexFilters` fields combine into the final query filter. Applies to both Vespa and OpenSearch.
+
+## Filter categories
+
+| Category | Fields | Join logic |
+|---|---|---|
+| **Visibility** | `hidden` | Always applied (unless `include_hidden`) |
+| **Tenant** | `tenant_id` | AND (multi-tenant only) |
+| **ACL** | `access_control_list` | OR within, AND with rest |
+| **Narrowing** | `source_type`, `tags`, `time_cutoff` | Each OR within, AND with rest |
+| **Knowledge scope** | `document_set`, `user_file_ids`, `attached_document_ids`, `hierarchy_node_ids` | OR within group, AND with rest |
+| **Additive scope** | `project_id`, `persona_id` | OR'd into knowledge scope **only when** a knowledge scope filter already exists |
+
+## How filters combine
+
+All categories are AND'd together. Within the knowledge scope category, individual filters are OR'd.
+
+```
+NOT hidden
+AND tenant = T                          -- if multi-tenant
+AND (acl contains A1 OR acl contains A2)
+AND (source_type = S1 OR ...)           -- if set
+AND (tag = T1 OR ...)                   -- if set
+AND <knowledge scope>                   -- see below
+AND time >= cutoff                      -- if set
+```
+
+## Knowledge scope rules
+
+The knowledge scope filter controls **what knowledge an assistant can access**.
+
+### No explicit knowledge attached
+
+When `document_set`, `user_file_ids`, `attached_document_ids`, and `hierarchy_node_ids` are all empty/None:
+
+- **No knowledge scope filter is applied.** The assistant can see everything (subject to ACL).
+- `project_id` and `persona_id` are ignored — they never restrict on their own.
+
+### One explicit knowledge type
+
+```
+-- Only document sets
+AND (document_sets contains "Engineering" OR document_sets contains "Legal")
+
+-- Only user files
+AND (document_id = "uuid-1" OR document_id = "uuid-2")
+```
+
+### Multiple explicit knowledge types (OR'd)
+
+```
+-- Document sets + user files
+AND (
+    document_sets contains "Engineering"
+    OR document_id = "uuid-1"
+)
+```
+
+### Explicit knowledge + overflowing user files
+
+When an explicit knowledge restriction is in effect **and** `project_id` or `persona_id` is set (user files overflowed the LLM context window), the additive scopes widen the filter:
+
+```
+-- Document sets + persona user files overflowed
+AND (
+    document_sets contains "Engineering"
+    OR personas contains 42
+)
+
+-- User files + project files overflowed
+AND (
+    document_id = "uuid-1"
+    OR user_project contains 7
+)
+```
+
+### Only project_id or persona_id (no explicit knowledge)
+
+No knowledge scope filter. The assistant searches everything.
+
+```
+-- Just ACL, no restriction
+NOT hidden
+AND (acl contains ...)
+```
+
+## Field reference
+
+| Filter field | Vespa field | Vespa type | Purpose |
+|---|---|---|---|
+| `document_set` | `document_sets` | `weightedset<string>` | Connector doc sets attached to assistant |
+| `user_file_ids` | `document_id` | `string` | User files uploaded to assistant |
+| `attached_document_ids` | `document_id` | `string` | Documents explicitly attached (OpenSearch only) |
+| `hierarchy_node_ids` | `ancestor_hierarchy_node_ids` | `array<int>` | Folder/space nodes (OpenSearch only) |
+| `project_id` | `user_project` | `array<int>` | Project tag for overflowing user files |
+| `persona_id` | `personas` | `array<int>` | Persona tag for overflowing user files |
+| `access_control_list` | `access_control_list` | `weightedset<string>` | ACL entries for the requesting user |
+| `source_type` | `source_type` | `string` | Connector source type (e.g. `web`, `jira`) |
+| `tags` | `metadata_list` | `array<string>` | Document metadata tags |
+| `time_cutoff` | `doc_updated_at` | `long` | Minimum document update timestamp |
+| `tenant_id` | `tenant_id` | `string` | Tenant isolation (multi-tenant) |

backend/onyx/document_index/opensearch/search.py

@@ -698,41 +698,6 @@ class DocumentQuery:
             """
             return {"terms": {ANCESTOR_HIERARCHY_NODE_IDS_FIELD_NAME: node_ids}}
 
-        def _get_assistant_knowledge_filter(
-            attached_doc_ids: list[str] | None,
-            node_ids: list[int] | None,
-            file_ids: list[UUID] | None,
-            document_sets: list[str] | None,
-        ) -> dict[str, Any]:
-            """Combined filter for assistant knowledge.
-
-            When an assistant has attached knowledge, search should be scoped to:
-            - Documents explicitly attached (by document ID), OR
-            - Documents under attached hierarchy nodes (by ancestor node IDs), OR
-            - User-uploaded files attached to the assistant, OR
-            - Documents in the assistant's document sets (if any)
-            """
-            knowledge_filter: dict[str, Any] = {
-                "bool": {"should": [], "minimum_should_match": 1}
-            }
-            if attached_doc_ids:
-                knowledge_filter["bool"]["should"].append(
-                    _get_attached_document_id_filter(attached_doc_ids)
-                )
-            if node_ids:
-                knowledge_filter["bool"]["should"].append(
-                    _get_hierarchy_node_filter(node_ids)
-                )
-            if file_ids:
-                knowledge_filter["bool"]["should"].append(
-                    _get_user_file_id_filter(file_ids)
-                )
-            if document_sets:
-                knowledge_filter["bool"]["should"].append(
-                    _get_document_set_filter(document_sets)
-                )
-            return knowledge_filter
-
         filter_clauses: list[dict[str, Any]] = []
 
         if not include_hidden:
@@ -758,41 +723,53 @@ class DocumentQuery:
             # document's metadata list.
             filter_clauses.append(_get_tag_filter(tags))
 
-        # Check if this is an assistant knowledge search (has any assistant-scoped knowledge)
-        has_assistant_knowledge = (
+        # Knowledge scope: explicit knowledge attachments restrict what
+        # an assistant can see.  When none are set the assistant
+        # searches everything.
+        #
+        # project_id / persona_id are additive: they make overflowing
+        # user files findable but must NOT trigger the restriction on
+        # their own (an agent with no explicit knowledge should search
+        # everything).
+        has_knowledge_scope = (
             attached_document_ids
             or hierarchy_node_ids
             or user_file_ids
             or document_sets
         )
 
-        if has_assistant_knowledge:
-            # If assistant has attached knowledge, scope search to that knowledge.
-            # Document sets are included in the OR filter so directly attached
-            # docs are always findable even if not in the document sets.
-            filter_clauses.append(
-                _get_assistant_knowledge_filter(
-                    attached_document_ids,
-                    hierarchy_node_ids,
-                    user_file_ids,
-                    document_sets,
+        if has_knowledge_scope:
+            knowledge_filter: dict[str, Any] = {
+                "bool": {"should": [], "minimum_should_match": 1}
+            }
+            if attached_document_ids:
+                knowledge_filter["bool"]["should"].append(
+                    _get_attached_document_id_filter(attached_document_ids)
                 )
-            )
-        elif user_file_ids:
-            # Fallback for non-assistant user file searches (e.g., project searches)
-            # If at least one user file ID is provided, the caller will only
-            # retrieve documents where the document ID is in this input list of
-            # file IDs.
-            filter_clauses.append(_get_user_file_id_filter(user_file_ids))
-
-        if project_id is not None:
-            # If a project ID is provided, the caller will only retrieve
-            # documents where the project ID provided here is present in the
-            # document's user projects list.
-            filter_clauses.append(_get_user_project_filter(project_id))
-
-        if persona_id is not None:
-            filter_clauses.append(_get_persona_filter(persona_id))
+            if hierarchy_node_ids:
+                knowledge_filter["bool"]["should"].append(
+                    _get_hierarchy_node_filter(hierarchy_node_ids)
+                )
+            if user_file_ids:
+                knowledge_filter["bool"]["should"].append(
+                    _get_user_file_id_filter(user_file_ids)
+                )
+            if document_sets:
+                knowledge_filter["bool"]["should"].append(
+                    _get_document_set_filter(document_sets)
+                )
+            # Additive: widen scope to also cover overflowing user
+            # files, but only when an explicit restriction is already
+            # in effect.
+            if project_id is not None:
+                knowledge_filter["bool"]["should"].append(
+                    _get_user_project_filter(project_id)
+                )
+            if persona_id is not None:
+                knowledge_filter["bool"]["should"].append(
+                    _get_persona_filter(persona_id)
+                )
+            filter_clauses.append(knowledge_filter)
 
         if time_cutoff is not None:
             # If a time cutoff is provided, the caller will only retrieve

backend/onyx/document_index/vespa/shared_utils/vespa_request_builders.py

@@ -23,11 +23,8 @@ from shared_configs.configs import MULTI_TENANT
 logger = setup_logger()
 
 
-def build_tenant_id_filter(tenant_id: str, include_trailing_and: bool = False) -> str:
-    filter_str = f'({TENANT_ID} contains "{tenant_id}")'
-    if include_trailing_and:
-        filter_str += " and "
-    return filter_str
+def build_tenant_id_filter(tenant_id: str) -> str:
+    return f'({TENANT_ID} contains "{tenant_id}")'
 
 
 def build_vespa_filters(
@@ -37,30 +34,22 @@ def build_vespa_filters(
     remove_trailing_and: bool = False,  # Set to True when using as a complete Vespa query
 ) -> str:
     def _build_or_filters(key: str, vals: list[str] | None) -> str:
-        """For string-based 'contains' filters, e.g. WSET fields or array<string> fields."""
+        """For string-based 'contains' filters, e.g. WSET fields or array<string> fields.
+        Returns a bare clause like '(key contains "v1" or key contains "v2")' or ""."""
         if not key or not vals:
             return ""
         eq_elems = [f'{key} contains "{val}"' for val in vals if val]
         if not eq_elems:
             return ""
-        or_clause = " or ".join(eq_elems)
-        return f"({or_clause}) and "
+        return f"({' or '.join(eq_elems)})"
 
     def _build_int_or_filters(key: str, vals: list[int] | None) -> str:
-        """
-        For an integer field filter.
-        If vals is not None, we want *only* docs whose key matches one of vals.
-        """
-        # If `vals` is None => skip the filter entirely
+        """For an integer field filter.
+        Returns a bare clause or ""."""
         if vals is None or not vals:
             return ""
-
-        # Otherwise build the OR filter
         eq_elems = [f"{key} = {val}" for val in vals]
-        or_clause = " or ".join(eq_elems)
-        result = f"({or_clause}) and "
-
-        return result
+        return f"({' or '.join(eq_elems)})"
 
     def _build_kg_filter(
         kg_entities: list[str] | None,
@@ -73,16 +62,12 @@ def build_vespa_filters(
         combined_filter_parts = []
 
         def _build_kge(entity: str) -> str:
-            # TYPE-SUBTYPE::ID -> "TYPE-SUBTYPE::ID"
-            # TYPE-SUBTYPE::*  -> ({prefix: true}"TYPE-SUBTYPE")
-            # TYPE::*          -> ({prefix: true}"TYPE")
             GENERAL = "::*"
             if entity.endswith(GENERAL):
                 return f'({{prefix: true}}"{entity.split(GENERAL, 1)[0]}")'
             else:
                 return f'"{entity}"'
 
-        # OR the entities (give new design)
         if kg_entities:
             filter_parts = []
             for kg_entity in kg_entities:
@@ -104,8 +89,7 @@ def build_vespa_filters(
 
         # TODO: remove kg terms entirely from prompts and codebase
 
-        # AND the combined filter parts
-        return f"({' and '.join(combined_filter_parts)}) and "
+        return f"({' and '.join(combined_filter_parts)})"
 
     def _build_kg_source_filters(
         kg_sources: list[str] | None,
@@ -114,16 +98,14 @@ def build_vespa_filters(
             return ""
 
         source_phrases = [f'{DOCUMENT_ID} contains "{source}"' for source in kg_sources]
-
-        return f"({' or '.join(source_phrases)}) and "
+        return f"({' or '.join(source_phrases)})"
 
     def _build_kg_chunk_id_zero_only_filter(
         kg_chunk_id_zero_only: bool,
     ) -> str:
         if not kg_chunk_id_zero_only:
             return ""
-
-        return "(chunk_id = 0 ) and "
+        return "(chunk_id = 0)"
 
     def _build_time_filter(
         cutoff: datetime | None,
@@ -135,8 +117,8 @@ def build_vespa_filters(
         cutoff_secs = int(cutoff.timestamp())
 
         if include_untimed:
-            return f"!({DOC_UPDATED_AT} < {cutoff_secs}) and "
-        return f"({DOC_UPDATED_AT} >= {cutoff_secs}) and "
+            return f"!({DOC_UPDATED_AT} < {cutoff_secs})"
+        return f"({DOC_UPDATED_AT} >= {cutoff_secs})"
 
     def _build_user_project_filter(
         project_id: int | None,
@@ -147,8 +129,7 @@ def build_vespa_filters(
             pid = int(project_id)
         except Exception:
             return ""
-        # Vespa YQL 'contains' expects a string literal; quote the integer
-        return f'({USER_PROJECT} contains "{pid}") and '
+        return f'({USER_PROJECT} contains "{pid}")'
 
     def _build_persona_filter(
         persona_id: int | None,
@@ -160,73 +141,94 @@ def build_vespa_filters(
         except Exception:
             logger.warning(f"Invalid persona ID: {persona_id}")
             return ""
-        return f'({PERSONAS} contains "{pid}") and '
+        return f'({PERSONAS} contains "{pid}")'
 
-    # Start building the filter string
-    filter_str = f"!({HIDDEN}=true) and " if not include_hidden else ""
+    def _append(parts: list[str], clause: str) -> None:
+        if clause:
+            parts.append(clause)
+
+    # Collect all top-level filter clauses, then join with " and " at the end.
+    filter_parts: list[str] = []
+
+    if not include_hidden:
+        filter_parts.append(f"!({HIDDEN}=true)")
 
     # TODO: add error condition if MULTI_TENANT and no tenant_id filter is set
-    # If running in multi-tenant mode
     if filters.tenant_id and MULTI_TENANT:
-        filter_str += build_tenant_id_filter(
-            filters.tenant_id, include_trailing_and=True
-        )
+        filter_parts.append(build_tenant_id_filter(filters.tenant_id))
 
     # ACL filters
     if filters.access_control_list is not None:
-        filter_str += _build_or_filters(
-            ACCESS_CONTROL_LIST, filters.access_control_list
+        _append(
+            filter_parts,
+            _build_or_filters(ACCESS_CONTROL_LIST, filters.access_control_list),
         )
 
     # Source type filters
     source_strs = (
         [s.value for s in filters.source_type] if filters.source_type else None
     )
-    filter_str += _build_or_filters(SOURCE_TYPE, source_strs)
+    _append(filter_parts, _build_or_filters(SOURCE_TYPE, source_strs))
 
     # Tag filters
     tag_attributes = None
     if filters.tags:
-        # build e.g. "tag_key|tag_value"
         tag_attributes = [
             f"{tag.tag_key}{INDEX_SEPARATOR}{tag.tag_value}" for tag in filters.tags
         ]
-    filter_str += _build_or_filters(METADATA_LIST, tag_attributes)
+    _append(filter_parts, _build_or_filters(METADATA_LIST, tag_attributes))
 
-    # Document sets
-    filter_str += _build_or_filters(DOCUMENT_SETS, filters.document_set)
+    # Knowledge scope: explicit knowledge attachments (document_sets,
+    # user_file_ids) restrict what an assistant can see.  When none are
+    # set, the assistant can see everything.
+    #
+    # project_id / persona_id are additive: they make overflowing user
+    # files findable in Vespa but must NOT trigger the restriction on
+    # their own (an agent with no explicit knowledge should search
+    # everything).
+    knowledge_scope_parts: list[str] = []
+
+    _append(
+        knowledge_scope_parts, _build_or_filters(DOCUMENT_SETS, filters.document_set)
+    )
 
-    # Convert UUIDs to strings for user_file_ids
     user_file_ids_str = (
         [str(uuid) for uuid in filters.user_file_ids] if filters.user_file_ids else None
     )
-    filter_str += _build_or_filters(DOCUMENT_ID, user_file_ids_str)
+    _append(knowledge_scope_parts, _build_or_filters(DOCUMENT_ID, user_file_ids_str))
 
-    # User project filter (array<int> attribute membership)
-    filter_str += _build_user_project_filter(filters.project_id)
+    # Only include project/persona scopes when an explicit knowledge
+    # restriction is already in effect — they widen the scope to also
+    # cover overflowing user files but never restrict on their own.
+    if knowledge_scope_parts:
+        _append(knowledge_scope_parts, _build_user_project_filter(filters.project_id))
+        _append(knowledge_scope_parts, _build_persona_filter(filters.persona_id))
 
-    # Persona filter (array<int> attribute membership)
-    filter_str += _build_persona_filter(filters.persona_id)
+    if len(knowledge_scope_parts) > 1:
+        filter_parts.append("(" + " or ".join(knowledge_scope_parts) + ")")
+    elif len(knowledge_scope_parts) == 1:
+        filter_parts.append(knowledge_scope_parts[0])
 
     # Time filter
-    filter_str += _build_time_filter(filters.time_cutoff)
+    _append(filter_parts, _build_time_filter(filters.time_cutoff))
 
     # # Knowledge Graph Filters
-    # filter_str += _build_kg_filter(
+    # _append(filter_parts, _build_kg_filter(
     #     kg_entities=filters.kg_entities,
     #     kg_relationships=filters.kg_relationships,
     #     kg_terms=filters.kg_terms,
-    # )
+    # ))
 
-    # filter_str += _build_kg_source_filters(filters.kg_sources)
+    # _append(filter_parts, _build_kg_source_filters(filters.kg_sources))
 
-    # filter_str += _build_kg_chunk_id_zero_only_filter(
+    # _append(filter_parts, _build_kg_chunk_id_zero_only_filter(
     #     filters.kg_chunk_id_zero_only or False
-    # )
+    # ))
 
-    # Trim trailing " and "
-    if remove_trailing_and and filter_str.endswith(" and "):
-        filter_str = filter_str[:-5]
+    filter_str = " and ".join(filter_parts)
+
+    if filter_str and not remove_trailing_and:
+        filter_str += " and "
 
     return filter_str
 

backend/onyx/utils/encryption.py

@@ -11,16 +11,18 @@ logger = setup_logger()
 
 
 # IMPORTANT DO NOT DELETE, THIS IS USED BY fetch_versioned_implementation
-def _encrypt_string(input_str: str) -> bytes:
-    if ENCRYPTION_KEY_SECRET:
+def _encrypt_string(input_str: str, key: str | None = None) -> bytes:
+    if ENCRYPTION_KEY_SECRET or key:
         logger.warning("MIT version of Onyx does not support encryption of secrets.")
     return input_str.encode()
 
 
 # IMPORTANT DO NOT DELETE, THIS IS USED BY fetch_versioned_implementation
-def _decrypt_bytes(input_bytes: bytes) -> str:
-    # No need to double warn. If you wish to learn more about encryption features
-    # refer to the Onyx EE code
+def _decrypt_bytes(input_bytes: bytes, key: str | None = None) -> str:  # noqa: ARG001
+    if key is not None:
+        logger.warning(
+            "MIT version of Onyx does not support decryption with a provided key."
+        )
     return input_bytes.decode()
 
 
@@ -86,15 +88,15 @@ def _mask_list(items: list[Any]) -> list[Any]:
     return masked
 
 
-def encrypt_string_to_bytes(intput_str: str) -> bytes:
+def encrypt_string_to_bytes(intput_str: str, key: str | None = None) -> bytes:
     versioned_encryption_fn = fetch_versioned_implementation(
         "onyx.utils.encryption", "_encrypt_string"
     )
-    return versioned_encryption_fn(intput_str)
+    return versioned_encryption_fn(intput_str, key=key)
 
 
-def decrypt_bytes_to_string(intput_bytes: bytes) -> str:
+def decrypt_bytes_to_string(intput_bytes: bytes, key: str | None = None) -> str:
     versioned_decryption_fn = fetch_versioned_implementation(
         "onyx.utils.encryption", "_decrypt_bytes"
     )
-    return versioned_decryption_fn(intput_bytes)
+    return versioned_decryption_fn(intput_bytes, key=key)

backend/onyx/utils/jsonriver/__init__.py

@@ -0,0 +1,17 @@
+"""
+jsonriver - A streaming JSON parser for Python
+
+Parse JSON incrementally as it streams in, e.g. from a network request or a language model.
+Gives you a sequence of increasingly complete values.
+
+Copyright (c) 2023 Google LLC (original TypeScript implementation)
+Copyright (c) 2024 jsonriver-python contributors (Python port)
+SPDX-License-Identifier: BSD-3-Clause
+"""
+
+from .parse import _Parser as Parser
+from .parse import JsonObject
+from .parse import JsonValue
+
+__all__ = ["Parser", "JsonValue", "JsonObject"]
+__version__ = "0.0.1"

backend/onyx/utils/jsonriver/parse.py

@@ -0,0 +1,427 @@
+"""
+JSON parser for streaming incremental parsing
+
+Copyright (c) 2023 Google LLC (original TypeScript implementation)
+Copyright (c) 2024 jsonriver-python contributors (Python port)
+SPDX-License-Identifier: BSD-3-Clause
+"""
+
+from __future__ import annotations
+
+import copy
+from enum import IntEnum
+from typing import cast
+from typing import Union
+
+from .tokenize import _Input
+from .tokenize import json_token_type_to_string
+from .tokenize import JsonTokenType
+from .tokenize import Tokenizer
+
+
+# Type definitions for JSON values
+JsonValue = Union[None, bool, float, str, list["JsonValue"], dict[str, "JsonValue"]]
+JsonObject = dict[str, JsonValue]
+
+
+class _StateEnum(IntEnum):
+    """Parser state machine states"""
+
+    Initial = 0
+    InString = 1
+    InArray = 2
+    InObjectExpectingKey = 3
+    InObjectExpectingValue = 4
+
+
+class _State:
+    """Base class for parser states"""
+
+    type: _StateEnum
+    value: JsonValue | tuple[str, JsonObject] | None
+
+
+class _InitialState(_State):
+    """Initial state before any parsing"""
+
+    def __init__(self) -> None:
+        self.type = _StateEnum.Initial
+        self.value = None
+
+
+class _InStringState(_State):
+    """State while parsing a string"""
+
+    def __init__(self) -> None:
+        self.type = _StateEnum.InString
+        self.value = ""
+
+
+class _InArrayState(_State):
+    """State while parsing an array"""
+
+    def __init__(self) -> None:
+        self.type = _StateEnum.InArray
+        self.value: list[JsonValue] = []
+
+
+class _InObjectExpectingKeyState(_State):
+    """State while parsing an object, expecting a key"""
+
+    def __init__(self) -> None:
+        self.type = _StateEnum.InObjectExpectingKey
+        self.value: JsonObject = {}
+
+
+class _InObjectExpectingValueState(_State):
+    """State while parsing an object, expecting a value"""
+
+    def __init__(self, key: str, obj: JsonObject) -> None:
+        self.type = _StateEnum.InObjectExpectingValue
+        self.value = (key, obj)
+
+
+# Sentinel value to distinguish "not set" from "set to None/null"
+class _Unset:
+    pass
+
+
+_UNSET = _Unset()
+
+
+class _Parser:
+    """
+    Incremental JSON parser
+
+    Feed chunks of JSON text via feed() and get back progressively
+    more complete JSON values.
+    """
+
+    def __init__(self) -> None:
+        self._state_stack: list[_State] = [_InitialState()]
+        self._toplevel_value: JsonValue | _Unset = _UNSET
+        self._input = _Input()
+        self.tokenizer = Tokenizer(self._input, self)
+        self._finished = False
+        self._progressed = False
+        self._prev_snapshot: JsonValue | _Unset = _UNSET
+
+    def feed(self, chunk: str) -> list[JsonValue]:
+        """
+        Feed a chunk of JSON text and return deltas from the previous state.
+
+        Each element in the returned list represents what changed since the
+        last yielded value. For dicts, only changed/new keys are included,
+        with string values containing only the newly appended characters.
+        """
+        if self._finished:
+            return []
+
+        self._input.feed(chunk)
+        return self._collect_deltas()
+
+    @staticmethod
+    def _compute_delta(prev: JsonValue | None, current: JsonValue) -> JsonValue | None:
+        if prev is None:
+            return current
+
+        if isinstance(current, dict) and isinstance(prev, dict):
+            result: JsonObject = {}
+            for key in current:
+                cur_val = current[key]
+                prev_val = prev.get(key)
+                if key not in prev:
+                    result[key] = cur_val
+                elif isinstance(cur_val, str) and isinstance(prev_val, str):
+                    if cur_val != prev_val:
+                        result[key] = cur_val[len(prev_val) :]
+                elif isinstance(cur_val, list) and isinstance(prev_val, list):
+                    if cur_val != prev_val:
+                        new_items = cur_val[len(prev_val) :]
+                        # check if the last existing element was updated
+                        if (
+                            prev_val
+                            and len(cur_val) >= len(prev_val)
+                            and cur_val[len(prev_val) - 1] != prev_val[-1]
+                        ):
+                            result[key] = [cur_val[len(prev_val) - 1]] + new_items
+                        elif new_items:
+                            result[key] = new_items
+                elif cur_val != prev_val:
+                    result[key] = cur_val
+            return result if result else None
+
+        if isinstance(current, str) and isinstance(prev, str):
+            delta = current[len(prev) :]
+            return delta if delta else None
+
+        if isinstance(current, list) and isinstance(prev, list):
+            if current != prev:
+                new_items = current[len(prev) :]
+                if (
+                    prev
+                    and len(current) >= len(prev)
+                    and current[len(prev) - 1] != prev[-1]
+                ):
+                    return [current[len(prev) - 1]] + new_items
+                return new_items if new_items else None
+            return None
+
+        if current != prev:
+            return current
+        return None
+
+    def finish(self) -> list[JsonValue]:
+        """Signal that no more chunks will be fed. Validates trailing content.
+
+        Returns any final deltas produced by flushing pending tokens (e.g.
+        numbers, which have no terminator and wait for more input).
+        """
+        self._input.mark_complete()
+        # Pump once more so the tokenizer can emit tokens that were waiting
+        # for more input (e.g. numbers need buffer_complete to finalize).
+        results = self._collect_deltas()
+        self._input.expect_end_of_content()
+        return results
+
+    def _collect_deltas(self) -> list[JsonValue]:
+        """Run one pump cycle and return any deltas produced."""
+        results: list[JsonValue] = []
+        while True:
+            self._progressed = False
+            self.tokenizer.pump()
+
+            if self._progressed:
+                if self._toplevel_value is _UNSET:
+                    raise RuntimeError(
+                        "Internal error: toplevel_value should not be unset "
+                        "after progressing"
+                    )
+                current = copy.deepcopy(cast(JsonValue, self._toplevel_value))
+                if isinstance(self._prev_snapshot, _Unset):
+                    results.append(current)
+                else:
+                    delta = self._compute_delta(self._prev_snapshot, current)
+                    if delta is not None:
+                        results.append(delta)
+                self._prev_snapshot = current
+            else:
+                if not self._state_stack:
+                    self._finished = True
+                break
+        return results
+
+    # TokenHandler protocol implementation
+
+    def handle_null(self) -> None:
+        """Handle null token"""
+        self._handle_value_token(JsonTokenType.Null, None)
+
+    def handle_boolean(self, value: bool) -> None:
+        """Handle boolean token"""
+        self._handle_value_token(JsonTokenType.Boolean, value)
+
+    def handle_number(self, value: float) -> None:
+        """Handle number token"""
+        self._handle_value_token(JsonTokenType.Number, value)
+
+    def handle_string_start(self) -> None:
+        """Handle string start token"""
+        state = self._current_state()
+        if not self._progressed and state.type != _StateEnum.InObjectExpectingKey:
+            self._progressed = True
+
+        if state.type == _StateEnum.Initial:
+            self._state_stack.pop()
+            self._toplevel_value = self._progress_value(JsonTokenType.StringStart, None)
+
+        elif state.type == _StateEnum.InArray:
+            v = self._progress_value(JsonTokenType.StringStart, None)
+            arr = cast(list[JsonValue], state.value)
+            arr.append(v)
+
+        elif state.type == _StateEnum.InObjectExpectingKey:
+            self._state_stack.append(_InStringState())
+
+        elif state.type == _StateEnum.InObjectExpectingValue:
+            key, obj = cast(tuple[str, JsonObject], state.value)
+            sv = self._progress_value(JsonTokenType.StringStart, None)
+            obj[key] = sv
+
+        elif state.type == _StateEnum.InString:
+            raise ValueError(
+                f"Unexpected {json_token_type_to_string(JsonTokenType.StringStart)} "
+                f"token in the middle of string"
+            )
+
+    def handle_string_middle(self, value: str) -> None:
+        """Handle string middle token"""
+        state = self._current_state()
+
+        if not self._progressed:
+            if len(self._state_stack) >= 2:
+                prev = self._state_stack[-2]
+                if prev.type != _StateEnum.InObjectExpectingKey:
+                    self._progressed = True
+            else:
+                self._progressed = True
+
+        if state.type != _StateEnum.InString:
+            raise ValueError(
+                f"Unexpected {json_token_type_to_string(JsonTokenType.StringMiddle)} "
+                f"token when not in string"
+            )
+
+        assert isinstance(state.value, str)
+        state.value += value
+
+        parent_state = self._state_stack[-2] if len(self._state_stack) >= 2 else None
+        self._update_string_parent(state.value, parent_state)
+
+    def handle_string_end(self) -> None:
+        """Handle string end token"""
+        state = self._current_state()
+
+        if state.type != _StateEnum.InString:
+            raise ValueError(
+                f"Unexpected {json_token_type_to_string(JsonTokenType.StringEnd)} "
+                f"token when not in string"
+            )
+
+        self._state_stack.pop()
+        parent_state = self._state_stack[-1] if self._state_stack else None
+        assert isinstance(state.value, str)
+        self._update_string_parent(state.value, parent_state)
+
+    def handle_array_start(self) -> None:
+        """Handle array start token"""
+        self._handle_value_token(JsonTokenType.ArrayStart, None)
+
+    def handle_array_end(self) -> None:
+        """Handle array end token"""
+        state = self._current_state()
+        if state.type != _StateEnum.InArray:
+            raise ValueError(
+                f"Unexpected {json_token_type_to_string(JsonTokenType.ArrayEnd)} token"
+            )
+        self._state_stack.pop()
+
+    def handle_object_start(self) -> None:
+        """Handle object start token"""
+        self._handle_value_token(JsonTokenType.ObjectStart, None)
+
+    def handle_object_end(self) -> None:
+        """Handle object end token"""
+        state = self._current_state()
+
+        if state.type in (
+            _StateEnum.InObjectExpectingKey,
+            _StateEnum.InObjectExpectingValue,
+        ):
+            self._state_stack.pop()
+        else:
+            raise ValueError(
+                f"Unexpected {json_token_type_to_string(JsonTokenType.ObjectEnd)} token"
+            )
+
+    # Private helper methods
+
+    def _current_state(self) -> _State:
+        """Get current parser state"""
+        if not self._state_stack:
+            raise ValueError("Unexpected trailing input")
+        return self._state_stack[-1]
+
+    def _handle_value_token(self, token_type: JsonTokenType, value: JsonValue) -> None:
+        """Handle a complete value token"""
+        state = self._current_state()
+
+        if not self._progressed:
+            self._progressed = True
+
+        if state.type == _StateEnum.Initial:
+            self._state_stack.pop()
+            self._toplevel_value = self._progress_value(token_type, value)
+
+        elif state.type == _StateEnum.InArray:
+            v = self._progress_value(token_type, value)
+            arr = cast(list[JsonValue], state.value)
+            arr.append(v)
+
+        elif state.type == _StateEnum.InObjectExpectingValue:
+            key, obj = cast(tuple[str, JsonObject], state.value)
+            if token_type != JsonTokenType.StringStart:
+                self._state_stack.pop()
+                new_state = _InObjectExpectingKeyState()
+                new_state.value = obj
+                self._state_stack.append(new_state)
+
+            v = self._progress_value(token_type, value)
+            obj[key] = v
+
+        elif state.type == _StateEnum.InString:
+            raise ValueError(
+                f"Unexpected {json_token_type_to_string(token_type)} "
+                f"token in the middle of string"
+            )
+
+        elif state.type == _StateEnum.InObjectExpectingKey:
+            raise ValueError(
+                f"Unexpected {json_token_type_to_string(token_type)} "
+                f"token in the middle of object expecting key"
+            )
+
+    def _update_string_parent(self, updated: str, parent_state: _State | None) -> None:
+        """Update parent container with updated string value"""
+        if parent_state is None:
+            self._toplevel_value = updated
+
+        elif parent_state.type == _StateEnum.InArray:
+            arr = cast(list[JsonValue], parent_state.value)
+            arr[-1] = updated
+
+        elif parent_state.type == _StateEnum.InObjectExpectingValue:
+            key, obj = cast(tuple[str, JsonObject], parent_state.value)
+            obj[key] = updated
+            if self._state_stack and self._state_stack[-1] == parent_state:
+                self._state_stack.pop()
+                new_state = _InObjectExpectingKeyState()
+                new_state.value = obj
+                self._state_stack.append(new_state)
+
+        elif parent_state.type == _StateEnum.InObjectExpectingKey:
+            if self._state_stack and self._state_stack[-1] == parent_state:
+                self._state_stack.pop()
+                obj = cast(JsonObject, parent_state.value)
+                self._state_stack.append(_InObjectExpectingValueState(updated, obj))
+
+    def _progress_value(self, token_type: JsonTokenType, value: JsonValue) -> JsonValue:
+        """Create initial value for a token and push appropriate state"""
+        if token_type == JsonTokenType.Null:
+            return None
+
+        elif token_type == JsonTokenType.Boolean:
+            return value
+
+        elif token_type == JsonTokenType.Number:
+            return value
+
+        elif token_type == JsonTokenType.StringStart:
+            string_state = _InStringState()
+            self._state_stack.append(string_state)
+            return ""
+
+        elif token_type == JsonTokenType.ArrayStart:
+            array_state = _InArrayState()
+            self._state_stack.append(array_state)
+            return array_state.value
+
+        elif token_type == JsonTokenType.ObjectStart:
+            object_state = _InObjectExpectingKeyState()
+            self._state_stack.append(object_state)
+            return object_state.value
+
+        else:
+            raise ValueError(
+                f"Unexpected token type: {json_token_type_to_string(token_type)}"
+            )

backend/onyx/utils/jsonriver/tokenize.py

@@ -0,0 +1,514 @@
+"""
+JSON tokenizer for streaming incremental parsing
+
+Copyright (c) 2023 Google LLC (original TypeScript implementation)
+Copyright (c) 2024 jsonriver-python contributors (Python port)
+SPDX-License-Identifier: BSD-3-Clause
+"""
+
+from __future__ import annotations
+
+import re
+from enum import IntEnum
+from typing import Protocol
+
+
+class TokenHandler(Protocol):
+    """Protocol for handling JSON tokens"""
+
+    def handle_null(self) -> None: ...
+    def handle_boolean(self, value: bool) -> None: ...
+    def handle_number(self, value: float) -> None: ...
+    def handle_string_start(self) -> None: ...
+    def handle_string_middle(self, value: str) -> None: ...
+    def handle_string_end(self) -> None: ...
+    def handle_array_start(self) -> None: ...
+    def handle_array_end(self) -> None: ...
+    def handle_object_start(self) -> None: ...
+    def handle_object_end(self) -> None: ...
+
+
+class JsonTokenType(IntEnum):
+    """Types of JSON tokens"""
+
+    Null = 0
+    Boolean = 1
+    Number = 2
+    StringStart = 3
+    StringMiddle = 4
+    StringEnd = 5
+    ArrayStart = 6
+    ArrayEnd = 7
+    ObjectStart = 8
+    ObjectEnd = 9
+
+
+def json_token_type_to_string(token_type: JsonTokenType) -> str:
+    """Convert token type to readable string"""
+    names = {
+        JsonTokenType.Null: "null",
+        JsonTokenType.Boolean: "boolean",
+        JsonTokenType.Number: "number",
+        JsonTokenType.StringStart: "string start",
+        JsonTokenType.StringMiddle: "string middle",
+        JsonTokenType.StringEnd: "string end",
+        JsonTokenType.ArrayStart: "array start",
+        JsonTokenType.ArrayEnd: "array end",
+        JsonTokenType.ObjectStart: "object start",
+        JsonTokenType.ObjectEnd: "object end",
+    }
+    return names[token_type]
+
+
+class _State(IntEnum):
+    """Internal tokenizer states"""
+
+    ExpectingValue = 0
+    InString = 1
+    StartArray = 2
+    AfterArrayValue = 3
+    StartObject = 4
+    AfterObjectKey = 5
+    AfterObjectValue = 6
+    BeforeObjectKey = 7
+
+
+# Regex for validating JSON numbers
+_JSON_NUMBER_PATTERN = re.compile(r"^-?(0|[1-9]\d*)(\.\d+)?([eE][+-]?\d+)?$")
+
+
+def _parse_json_number(s: str) -> float:
+    """Parse a JSON number string, validating format"""
+    if not _JSON_NUMBER_PATTERN.match(s):
+        raise ValueError("Invalid number")
+    return float(s)
+
+
+class _Input:
+    """
+    Input buffer for chunk-based JSON parsing
+
+    Manages buffering of input chunks and provides methods for
+    consuming and inspecting the buffer.
+    """
+
+    def __init__(self) -> None:
+        self._buffer = ""
+        self._start_index = 0
+        self.buffer_complete = False
+
+    def feed(self, chunk: str) -> None:
+        """Add a chunk of data to the buffer"""
+        self._buffer += chunk
+
+    def mark_complete(self) -> None:
+        """Signal that no more chunks will be fed"""
+        self.buffer_complete = True
+
+    @property
+    def length(self) -> int:
+        """Number of characters remaining in buffer"""
+        return len(self._buffer) - self._start_index
+
+    def advance(self, length: int) -> None:
+        """Advance the start position by length characters"""
+        self._start_index += length
+
+    def peek(self, offset: int) -> str | None:
+        """Peek at character at offset, or None if not available"""
+        idx = self._start_index + offset
+        if idx < len(self._buffer):
+            return self._buffer[idx]
+        return None
+
+    def peek_char_code(self, offset: int) -> int:
+        """Get character code at offset"""
+        return ord(self._buffer[self._start_index + offset])
+
+    def slice(self, start: int, end: int) -> str:
+        """Slice buffer from start to end (relative to current position)"""
+        return self._buffer[self._start_index + start : self._start_index + end]
+
+    def commit(self) -> None:
+        """Commit consumed content, removing it from buffer"""
+        if self._start_index > 0:
+            self._buffer = self._buffer[self._start_index :]
+            self._start_index = 0
+
+    def remaining(self) -> str:
+        """Get all remaining content in buffer"""
+        return self._buffer[self._start_index :]
+
+    def expect_end_of_content(self) -> None:
+        """Verify no non-whitespace content remains"""
+        self.commit()
+        self.skip_past_whitespace()
+        if self.length != 0:
+            raise ValueError(f"Unexpected trailing content {self.remaining()!r}")
+
+    def skip_past_whitespace(self) -> None:
+        """Skip whitespace characters"""
+        i = self._start_index
+        while i < len(self._buffer):
+            c = ord(self._buffer[i])
+            if c in (32, 9, 10, 13):  # space, tab, \n, \r
+                i += 1
+            else:
+                break
+        self._start_index = i
+
+    def try_to_take_prefix(self, prefix: str) -> bool:
+        """Try to consume prefix from buffer, return True if successful"""
+        if self._buffer.startswith(prefix, self._start_index):
+            self._start_index += len(prefix)
+            return True
+        return False
+
+    def try_to_take(self, length: int) -> str | None:
+        """Try to take length characters, or None if not enough available"""
+        if self.length < length:
+            return None
+        result = self._buffer[self._start_index : self._start_index + length]
+        self._start_index += length
+        return result
+
+    def try_to_take_char_code(self) -> int | None:
+        """Try to take a single character as char code, or None if buffer empty"""
+        if self.length == 0:
+            return None
+        code = ord(self._buffer[self._start_index])
+        self._start_index += 1
+        return code
+
+    def take_until_quote_or_backslash(self) -> tuple[str, bool]:
+        """
+        Consume input up to first quote or backslash
+
+        Returns tuple of (consumed_content, pattern_found)
+        """
+        buf = self._buffer
+        i = self._start_index
+        while i < len(buf):
+            c = ord(buf[i])
+            if c <= 0x1F:
+                raise ValueError("Unescaped control character in string")
+            if c == 34 or c == 92:  # " or \
+                result = buf[self._start_index : i]
+                self._start_index = i
+                return (result, True)
+            i += 1
+
+        result = buf[self._start_index :]
+        self._start_index = len(buf)
+        return (result, False)
+
+
+class Tokenizer:
+    """
+    Tokenizer for chunk-based JSON parsing
+
+    Processes chunks fed into its input buffer and calls handler methods
+    as JSON tokens are recognized.
+    """
+
+    def __init__(self, input: _Input, handler: TokenHandler) -> None:
+        self.input = input
+        self._handler = handler
+        self._stack: list[_State] = [_State.ExpectingValue]
+        self._emitted_tokens = 0
+
+    def is_done(self) -> bool:
+        """Check if tokenization is complete"""
+        return len(self._stack) == 0 and self.input.length == 0
+
+    def pump(self) -> None:
+        """Process all available tokens in the buffer"""
+        while True:
+            before = self._emitted_tokens
+            self._tokenize_more()
+            if self._emitted_tokens == before:
+                self.input.commit()
+                return
+
+    def _tokenize_more(self) -> None:
+        """Process one step of tokenization based on current state"""
+        if not self._stack:
+            return
+
+        state = self._stack[-1]
+
+        if state == _State.ExpectingValue:
+            self._tokenize_value()
+        elif state == _State.InString:
+            self._tokenize_string()
+        elif state == _State.StartArray:
+            self._tokenize_array_start()
+        elif state == _State.AfterArrayValue:
+            self._tokenize_after_array_value()
+        elif state == _State.StartObject:
+            self._tokenize_object_start()
+        elif state == _State.AfterObjectKey:
+            self._tokenize_after_object_key()
+        elif state == _State.AfterObjectValue:
+            self._tokenize_after_object_value()
+        elif state == _State.BeforeObjectKey:
+            self._tokenize_before_object_key()
+
+    def _tokenize_value(self) -> None:
+        """Tokenize a JSON value"""
+        self.input.skip_past_whitespace()
+
+        if self.input.try_to_take_prefix("null"):
+            self._handler.handle_null()
+            self._emitted_tokens += 1
+            self._stack.pop()
+            return
+
+        if self.input.try_to_take_prefix("true"):
+            self._handler.handle_boolean(True)
+            self._emitted_tokens += 1
+            self._stack.pop()
+            return
+
+        if self.input.try_to_take_prefix("false"):
+            self._handler.handle_boolean(False)
+            self._emitted_tokens += 1
+            self._stack.pop()
+            return
+
+        if self.input.length > 0:
+            ch = self.input.peek_char_code(0)
+            if (48 <= ch <= 57) or ch == 45:  # 0-9 or -
+                # Scan for end of number
+                i = 0
+                while i < self.input.length:
+                    c = self.input.peek_char_code(i)
+                    if (48 <= c <= 57) or c in (45, 43, 46, 101, 69):  # 0-9 - + . e E
+                        i += 1
+                    else:
+                        break
+
+                if i == self.input.length and not self.input.buffer_complete:
+                    # Need more input (numbers have no terminator)
+                    return
+
+                number_chars = self.input.slice(0, i)
+                self.input.advance(i)
+                number = _parse_json_number(number_chars)
+                self._handler.handle_number(number)
+                self._emitted_tokens += 1
+                self._stack.pop()
+                return
+
+        if self.input.try_to_take_prefix('"'):
+            self._stack.pop()
+            self._stack.append(_State.InString)
+            self._handler.handle_string_start()
+            self._emitted_tokens += 1
+            self._tokenize_string()
+            return
+
+        if self.input.try_to_take_prefix("["):
+            self._stack.pop()
+            self._stack.append(_State.StartArray)
+            self._handler.handle_array_start()
+            self._emitted_tokens += 1
+            self._tokenize_array_start()
+            return
+
+        if self.input.try_to_take_prefix("{"):
+            self._stack.pop()
+            self._stack.append(_State.StartObject)
+            self._handler.handle_object_start()
+            self._emitted_tokens += 1
+            self._tokenize_object_start()
+            return
+
+    def _tokenize_string(self) -> None:
+        """Tokenize string content"""
+        while True:
+            chunk, interrupted = self.input.take_until_quote_or_backslash()
+            if chunk:
+                self._handler.handle_string_middle(chunk)
+                self._emitted_tokens += 1
+            elif not interrupted:
+                return
+
+            if interrupted:
+                if self.input.length == 0:
+                    return
+
+                next_char = self.input.peek(0)
+                if next_char == '"':
+                    self.input.advance(1)
+                    self._handler.handle_string_end()
+                    self._emitted_tokens += 1
+                    self._stack.pop()
+                    return
+
+                # Handle escape sequences
+                next_char2 = self.input.peek(1)
+                if next_char2 is None:
+                    return
+
+                value: str
+                if next_char2 == "u":
+                    # Unicode escape: need 4 hex digits
+                    if self.input.length < 6:
+                        return
+
+                    code = 0
+                    for j in range(2, 6):
+                        c = self.input.peek_char_code(j)
+                        if 48 <= c <= 57:  # 0-9
+                            digit = c - 48
+                        elif 65 <= c <= 70:  # A-F
+                            digit = c - 55
+                        elif 97 <= c <= 102:  # a-f
+                            digit = c - 87
+                        else:
+                            raise ValueError("Bad Unicode escape in JSON")
+                        code = (code << 4) | digit
+
+                    self.input.advance(6)
+                    self._handler.handle_string_middle(chr(code))
+                    self._emitted_tokens += 1
+                    continue
+
+                elif next_char2 == "n":
+                    value = "\n"
+                elif next_char2 == "r":
+                    value = "\r"
+                elif next_char2 == "t":
+                    value = "\t"
+                elif next_char2 == "b":
+                    value = "\b"
+                elif next_char2 == "f":
+                    value = "\f"
+                elif next_char2 == "\\":
+                    value = "\\"
+                elif next_char2 == "/":
+                    value = "/"
+                elif next_char2 == '"':
+                    value = '"'
+                else:
+                    raise ValueError("Bad escape in string")
+
+                self.input.advance(2)
+                self._handler.handle_string_middle(value)
+                self._emitted_tokens += 1
+
+    def _tokenize_array_start(self) -> None:
+        """Tokenize start of array (check for empty or first element)"""
+        self.input.skip_past_whitespace()
+        if self.input.length == 0:
+            return
+
+        if self.input.try_to_take_prefix("]"):
+            self._handler.handle_array_end()
+            self._emitted_tokens += 1
+            self._stack.pop()
+            return
+
+        self._stack.pop()
+        self._stack.append(_State.AfterArrayValue)
+        self._stack.append(_State.ExpectingValue)
+        self._tokenize_value()
+
+    def _tokenize_after_array_value(self) -> None:
+        """Tokenize after an array value (expect , or ])"""
+        self.input.skip_past_whitespace()
+        next_char = self.input.try_to_take_char_code()
+
+        if next_char is None:
+            return
+        elif next_char == 0x5D:  # ]
+            self._handler.handle_array_end()
+            self._emitted_tokens += 1
+            self._stack.pop()
+            return
+        elif next_char == 0x2C:  # ,
+            self._stack.append(_State.ExpectingValue)
+            self._tokenize_value()
+            return
+        else:
+            raise ValueError(f"Expected , or ], got {chr(next_char)!r}")
+
+    def _tokenize_object_start(self) -> None:
+        """Tokenize start of object (check for empty or first key)"""
+        self.input.skip_past_whitespace()
+        next_char = self.input.try_to_take_char_code()
+
+        if next_char is None:
+            return
+        elif next_char == 0x7D:  # }
+            self._handler.handle_object_end()
+            self._emitted_tokens += 1
+            self._stack.pop()
+            return
+        elif next_char == 0x22:  # "
+            self._stack.pop()
+            self._stack.append(_State.AfterObjectKey)
+            self._stack.append(_State.InString)
+            self._handler.handle_string_start()
+            self._emitted_tokens += 1
+            self._tokenize_string()
+            return
+        else:
+            raise ValueError(f"Expected start of object key, got {chr(next_char)!r}")
+
+    def _tokenize_after_object_key(self) -> None:
+        """Tokenize after object key (expect :)"""
+        self.input.skip_past_whitespace()
+        next_char = self.input.try_to_take_char_code()
+
+        if next_char is None:
+            return
+        elif next_char == 0x3A:  # :
+            self._stack.pop()
+            self._stack.append(_State.AfterObjectValue)
+            self._stack.append(_State.ExpectingValue)
+            self._tokenize_value()
+            return
+        else:
+            raise ValueError(f"Expected colon after object key, got {chr(next_char)!r}")
+
+    def _tokenize_after_object_value(self) -> None:
+        """Tokenize after object value (expect , or })"""
+        self.input.skip_past_whitespace()
+        next_char = self.input.try_to_take_char_code()
+
+        if next_char is None:
+            return
+        elif next_char == 0x7D:  # }
+            self._handler.handle_object_end()
+            self._emitted_tokens += 1
+            self._stack.pop()
+            return
+        elif next_char == 0x2C:  # ,
+            self._stack.pop()
+            self._stack.append(_State.BeforeObjectKey)
+            self._tokenize_before_object_key()
+            return
+        else:
+            raise ValueError(
+                f"Expected , or }} after object value, got {chr(next_char)!r}"
+            )
+
+    def _tokenize_before_object_key(self) -> None:
+        """Tokenize before object key (after comma)"""
+        self.input.skip_past_whitespace()
+        next_char = self.input.try_to_take_char_code()
+
+        if next_char is None:
+            return
+        elif next_char == 0x22:  # "
+            self._stack.pop()
+            self._stack.append(_State.AfterObjectKey)
+            self._stack.append(_State.InString)
+            self._handler.handle_string_start()
+            self._emitted_tokens += 1
+            self._tokenize_string()
+            return
+        else:
+            raise ValueError(f"Expected start of object key, got {chr(next_char)!r}")

backend/onyx/utils/variable_functionality.py

@@ -24,6 +24,9 @@ class OnyxVersion:
     def set_ee(self) -> None:
         self._is_ee = True
 
+    def unset_ee(self) -> None:
+        self._is_ee = False
+
     def is_ee_version(self) -> bool:
         return self._is_ee
 

backend/scripts/reencrypt_secrets.py

@@ -0,0 +1,99 @@
+"""Re-encrypt secrets under the current ENCRYPTION_KEY_SECRET.
+
+Decrypts all encrypted columns using the old key (or raw decode if the old key
+is empty), then re-encrypts them with the current ENCRYPTION_KEY_SECRET.
+
+Usage (docker):
+    docker exec -it onyx-api_server-1 \
+        python -m scripts.reencrypt_secrets --old-key "previous-key"
+
+Usage (kubernetes):
+    kubectl exec -it <pod> -- \
+        python -m scripts.reencrypt_secrets --old-key "previous-key"
+
+Omit --old-key (or pass "") if secrets were not previously encrypted.
+
+For multi-tenant deployments, pass --tenant-id to target a specific tenant,
+or --all-tenants to iterate every tenant.
+"""
+
+import argparse
+import os
+import sys
+
+parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.append(parent_dir)
+
+from onyx.db.rotate_encryption_key import rotate_encryption_key  # noqa: E402
+from onyx.db.engine.sql_engine import get_session_with_tenant  # noqa: E402
+from onyx.db.engine.sql_engine import SqlEngine  # noqa: E402
+from onyx.db.engine.tenant_utils import get_all_tenant_ids  # noqa: E402
+from onyx.utils.variable_functionality import global_version  # noqa: E402
+from shared_configs.configs import POSTGRES_DEFAULT_SCHEMA  # noqa: E402
+
+
+def _run_for_tenant(tenant_id: str, old_key: str | None, dry_run: bool = False) -> None:
+    print(f"Re-encrypting secrets for tenant: {tenant_id}")
+    with get_session_with_tenant(tenant_id=tenant_id) as db_session:
+        results = rotate_encryption_key(db_session, old_key=old_key, dry_run=dry_run)
+
+    if results:
+        for col, count in results.items():
+            print(
+                f"  {col}: {count} row(s) {'would be ' if dry_run else ''}re-encrypted"
+            )
+    else:
+        print("No rows needed re-encryption.")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Re-encrypt secrets under the current encryption key."
+    )
+    parser.add_argument(
+        "--old-key",
+        default=None,
+        help="Previous encryption key. Omit or pass empty string if not applicable.",
+    )
+    parser.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="Show what would be re-encrypted without making changes.",
+    )
+
+    tenant_group = parser.add_mutually_exclusive_group()
+    tenant_group.add_argument(
+        "--tenant-id",
+        default=None,
+        help="Target a specific tenant schema.",
+    )
+    tenant_group.add_argument(
+        "--all-tenants",
+        action="store_true",
+        help="Iterate all tenants.",
+    )
+
+    args = parser.parse_args()
+
+    old_key = args.old_key if args.old_key else None
+
+    global_version.set_ee()
+    SqlEngine.init_engine(pool_size=5, max_overflow=2)
+
+    if args.dry_run:
+        print("DRY RUN — no changes will be made")
+
+    if args.all_tenants:
+        tenant_ids = get_all_tenant_ids()
+        print(f"Found {len(tenant_ids)} tenant(s)")
+        for tid in tenant_ids:
+            _run_for_tenant(tid, old_key, dry_run=args.dry_run)
+    else:
+        tenant_id = args.tenant_id or POSTGRES_DEFAULT_SCHEMA
+        _run_for_tenant(tenant_id, old_key, dry_run=args.dry_run)
+
+    print("Done.")
+
+
+if __name__ == "__main__":
+    main()

backend/tests/README.md

@@ -0,0 +1,71 @@
+# Backend Tests
+
+## Test Types
+
+There are four test categories, ordered by increasing scope:
+
+### Unit Tests (`tests/unit/`)
+
+No external services. Mock all I/O with `unittest.mock`. Use for complex, isolated
+logic (e.g. citation processing, encryption).
+
+```bash
+pytest -xv backend/tests/unit
+```
+
+### External Dependency Unit Tests (`tests/external_dependency_unit/`)
+
+External services (Postgres, Redis, Vespa, OpenAI, etc.) are running, but Onyx
+application containers are not. Tests call functions directly and can mock selectively.
+
+Use when you need a real database or real API calls but want control over setup.
+
+```bash
+python -m dotenv -f .vscode/.env run -- pytest backend/tests/external_dependency_unit
+```
+
+### Integration Tests (`tests/integration/`)
+
+Full Onyx deployment running. No mocking. Prefer this over other test types when possible.
+
+```bash
+python -m dotenv -f .vscode/.env run -- pytest backend/tests/integration
+```
+
+### Playwright / E2E Tests (`web/tests/e2e/`)
+
+Full stack including web server. Use for frontend-backend coordination.
+
+```bash
+npx playwright test <TEST_NAME>
+```
+
+## Shared Fixtures
+
+Shared fixtures live in `backend/tests/conftest.py`. Test subdirectories can define
+their own `conftest.py` for directory-scoped fixtures.
+
+## Best Practices
+
+### Use `enable_ee` fixture instead of inlining
+
+Enables EE mode for a test, with proper teardown and cache clearing.
+
+```python
+# Whole file (in a test module, NOT in conftest.py)
+pytestmark = pytest.mark.usefixtures("enable_ee")
+
+# Whole directory — add an autouse wrapper to the directory's conftest.py
+@pytest.fixture(autouse=True)
+def _enable_ee_for_directory(enable_ee: None) -> None:  # noqa: ARG001
+    """Wraps the shared enable_ee fixture with autouse for this directory."""
+
+# Single test
+def test_something(enable_ee: None) -> None: ...
+```
+
+**Note:** `pytestmark` in a `conftest.py` does NOT apply markers to tests in that
+directory — it only affects tests defined in the conftest itself (which is none).
+Use the autouse fixture wrapper pattern shown above instead.
+
+Do NOT inline `global_version.set_ee()` — always use the fixture.

backend/tests/conftest.py

@@ -0,0 +1,24 @@
+"""Root conftest — shared fixtures available to all test directories."""
+
+from collections.abc import Generator
+
+import pytest
+
+from onyx.utils.variable_functionality import fetch_versioned_implementation
+from onyx.utils.variable_functionality import global_version
+
+
+@pytest.fixture()
+def enable_ee() -> Generator[None, None, None]:
+    """Temporarily enable EE mode for a single test.
+
+    Restores the previous EE state and clears the versioned-implementation
+    cache on teardown so state doesn't leak between tests.
+    """
+    was_ee = global_version.is_ee_version()
+    global_version.set_ee()
+    fetch_versioned_implementation.cache_clear()
+    yield
+    if not was_ee:
+        global_version.unset_ee()
+    fetch_versioned_implementation.cache_clear()

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

@@ -45,7 +45,7 @@ def confluence_connector() -> ConfluenceConnector:
 def test_confluence_connector_permissions(
     mock_get_api_key: MagicMock,  # noqa: ARG001
     confluence_connector: ConfluenceConnector,
-    set_ee_on: None,  # noqa: ARG001
+    enable_ee: None,  # noqa: ARG001
 ) -> None:
     # Get all doc IDs from the full connector
     all_full_doc_ids = set()
@@ -93,7 +93,7 @@ def test_confluence_connector_permissions(
 def test_confluence_connector_restriction_handling(
     mock_get_api_key: MagicMock,  # noqa: ARG001
     mock_db_provider_class: MagicMock,
-    set_ee_on: None,  # noqa: ARG001
+    enable_ee: None,  # noqa: ARG001
 ) -> None:
     # Test space key
     test_space_key = "DailyPermS"

backend/tests/daily/connectors/conftest.py

@@ -4,8 +4,6 @@ from unittest.mock import patch
 
 import pytest
 
-from onyx.utils.variable_functionality import global_version
-
 
 @pytest.fixture
 def mock_get_unstructured_api_key() -> Generator[MagicMock, None, None]:
@@ -14,14 +12,3 @@ def mock_get_unstructured_api_key() -> Generator[MagicMock, None, None]:
         return_value=None,
     ) as mock:
         yield mock
-
-
-@pytest.fixture
-def set_ee_on() -> Generator[None, None, None]:
-    """Need EE to be enabled for these tests to work since
-    perm syncing is a an EE-only feature."""
-    global_version.set_ee()
-
-    yield
-
-    global_version._is_ee = False

backend/tests/daily/connectors/google_drive/test_drive_perm_sync.py

@@ -98,7 +98,7 @@ def _build_connector(
 
 def test_gdrive_perm_sync_with_real_data(
     google_drive_service_acct_connector_factory: Callable[..., GoogleDriveConnector],
-    set_ee_on: None,  # noqa: ARG001
+    enable_ee: None,  # noqa: ARG001
 ) -> None:
     """
     Test gdrive_doc_sync and gdrive_group_sync with real data from the test drive.

backend/tests/daily/connectors/slack/test_slack_perm_sync.py

@@ -1,12 +1,10 @@
 import time
-from collections.abc import Generator
 
 import pytest
 
 from onyx.connectors.models import HierarchyNode
 from onyx.connectors.models import SlimDocument
 from onyx.connectors.slack.connector import SlackConnector
-from onyx.utils.variable_functionality import global_version
 from tests.daily.connectors.utils import load_all_from_connector
 
 
@@ -19,16 +17,7 @@ PRIVATE_CHANNEL_USERS = [
     "test_user_2@onyx-test.com",
 ]
 
-
-@pytest.fixture(autouse=True)
-def set_ee_on() -> Generator[None, None, None]:
-    """Need EE to be enabled for these tests to work since
-    perm syncing is a an EE-only feature."""
-    global_version.set_ee()
-
-    yield
-
-    global_version._is_ee = False
+pytestmark = pytest.mark.usefixtures("enable_ee")
 
 
 @pytest.mark.parametrize(

backend/tests/daily/connectors/teams/test_teams_connector.py

@@ -1,13 +1,11 @@
 import os
 import time
-from collections.abc import Generator
 
 import pytest
 
 from onyx.access.models import ExternalAccess
 from onyx.connectors.models import HierarchyNode
 from onyx.connectors.teams.connector import TeamsConnector
-from onyx.utils.variable_functionality import global_version
 from tests.daily.connectors.teams.models import TeamsThread
 from tests.daily.connectors.utils import load_all_from_connector
 
@@ -168,18 +166,9 @@ def test_slim_docs_retrieval_from_teams_connector(
         _assert_is_valid_external_access(external_access=slim_doc.external_access)
 
 
-@pytest.fixture(autouse=False)
-def set_ee_on() -> Generator[None, None, None]:
-    """Need EE to be enabled for perm sync tests to work since
-    perm syncing is an EE-only feature."""
-    global_version.set_ee()
-    yield
-    global_version._is_ee = False
-
-
 def test_load_from_checkpoint_with_perm_sync(
     teams_connector: TeamsConnector,
-    set_ee_on: None,  # noqa: ARG001
+    enable_ee: None,  # noqa: ARG001
 ) -> None:
     """Test that load_from_checkpoint_with_perm_sync returns documents with external_access.
 

backend/tests/external_dependency_unit/connectors/jira/test_jira_doc_sync.py

@@ -1,5 +1,6 @@
 from typing import Any
 
+import pytest
 from pydantic import BaseModel
 from sqlalchemy.orm import Session
 
@@ -14,13 +15,14 @@ from onyx.db.models import ConnectorCredentialPair
 from onyx.db.models import Credential
 from onyx.db.utils import DocumentRow
 from onyx.db.utils import SortOrder
-from onyx.utils.variable_functionality import global_version
 
 
 # In order to get these tests to run, use the credentials from Bitwarden.
 # Search up "ENV vars for local and Github tests", and find the Jira relevant key-value pairs.
 # Required env vars: JIRA_USER_EMAIL, JIRA_API_TOKEN
 
+pytestmark = pytest.mark.usefixtures("enable_ee")
+
 
 class DocExternalAccessSet(BaseModel):
     """A version of DocExternalAccess that uses sets for comparison."""
@@ -52,9 +54,6 @@ def test_jira_doc_sync(
     This test uses the AS project which has applicationRole permission,
     meaning all documents should be marked as public.
     """
-    # NOTE: must set EE on or else the connector will skip the perm syncing
-    global_version.set_ee()
-
     try:
         # Use AS project specifically for this test
         connector_config = {
@@ -150,9 +149,6 @@ def test_jira_doc_sync_with_specific_permissions(
     This test uses a project that has specific user permissions to verify
     that specific users are correctly extracted.
     """
-    # NOTE: must set EE on or else the connector will skip the perm syncing
-    global_version.set_ee()
-
     try:
         # Use SUP project which has specific user permissions
         connector_config = {

backend/tests/external_dependency_unit/connectors/jira/test_jira_group_sync.py

@@ -1,5 +1,6 @@
 from typing import Any
 
+import pytest
 from sqlalchemy.orm import Session
 
 from ee.onyx.external_permissions.jira.group_sync import jira_group_sync
@@ -18,6 +19,8 @@ from tests.daily.connectors.confluence.models import ExternalUserGroupSet
 # Search up "ENV vars for local and Github tests", and find the Jira relevant key-value pairs.
 # Required env vars: JIRA_USER_EMAIL, JIRA_API_TOKEN
 
+pytestmark = pytest.mark.usefixtures("enable_ee")
+
 # Expected groups from the danswerai.atlassian.net Jira instance
 # Note: These groups are shared with Confluence since they're both Atlassian products
 # App accounts (bots, integrations) are filtered out

backend/tests/external_dependency_unit/db/test_rotate_encryption_key.py

@@ -0,0 +1,305 @@
+"""Tests for rotate_encryption_key against real Postgres.
+
+Uses real ORM models (Credential, InternetSearchProvider) and the actual
+Postgres database. Discovery is mocked in rotation tests to scope mutations
+to only the test rows — the real _discover_encrypted_columns walk is tested
+separately in TestDiscoverEncryptedColumns.
+
+Requires a running Postgres instance. Run with::
+
+    python -m dotenv -f .vscode/.env run -- pytest tests/external_dependency_unit/db/test_rotate_encryption_key.py
+"""
+
+import json
+from collections.abc import Generator
+from unittest.mock import patch
+
+import pytest
+from sqlalchemy import LargeBinary
+from sqlalchemy import select
+from sqlalchemy import text
+from sqlalchemy.orm import Session
+
+from ee.onyx.utils.encryption import _decrypt_bytes
+from ee.onyx.utils.encryption import _encrypt_string
+from ee.onyx.utils.encryption import _get_trimmed_key
+from onyx.configs.constants import DocumentSource
+from onyx.db.models import Credential
+from onyx.db.models import EncryptedJson
+from onyx.db.models import EncryptedString
+from onyx.db.models import InternetSearchProvider
+from onyx.db.rotate_encryption_key import _discover_encrypted_columns
+from onyx.db.rotate_encryption_key import rotate_encryption_key
+from onyx.utils.variable_functionality import fetch_versioned_implementation
+from onyx.utils.variable_functionality import global_version
+
+EE_MODULE = "ee.onyx.utils.encryption"
+ROTATE_MODULE = "onyx.db.rotate_encryption_key"
+
+OLD_KEY = "o" * 16
+NEW_KEY = "n" * 16
+
+
+@pytest.fixture(autouse=True)
+def _enable_ee() -> Generator[None, None, None]:
+    prev = global_version._is_ee
+    global_version.set_ee()
+    fetch_versioned_implementation.cache_clear()
+    yield
+    global_version._is_ee = prev
+    fetch_versioned_implementation.cache_clear()
+
+
+@pytest.fixture(autouse=True)
+def _clear_key_cache() -> None:
+    _get_trimmed_key.cache_clear()
+
+
+def _raw_credential_bytes(db_session: Session, credential_id: int) -> bytes | None:
+    """Read raw bytes from credential_json, bypassing the TypeDecorator."""
+    col = Credential.__table__.c.credential_json
+    stmt = select(col.cast(LargeBinary)).where(
+        Credential.__table__.c.id == credential_id
+    )
+    return db_session.execute(stmt).scalar()
+
+
+def _raw_isp_bytes(db_session: Session, isp_id: int) -> bytes | None:
+    """Read raw bytes from InternetSearchProvider.api_key."""
+    col = InternetSearchProvider.__table__.c.api_key
+    stmt = select(col.cast(LargeBinary)).where(
+        InternetSearchProvider.__table__.c.id == isp_id
+    )
+    return db_session.execute(stmt).scalar()
+
+
+class TestDiscoverEncryptedColumns:
+    """Verify _discover_encrypted_columns finds real production models."""
+
+    def test_discovers_credential_json(self) -> None:
+        results = _discover_encrypted_columns()
+        found = {
+            (model_cls.__tablename__, col_name, is_json)  # type: ignore[attr-defined]
+            for model_cls, col_name, _, is_json in results
+        }
+        assert ("credential", "credential_json", True) in found
+
+    def test_discovers_internet_search_provider_api_key(self) -> None:
+        results = _discover_encrypted_columns()
+        found = {
+            (model_cls.__tablename__, col_name, is_json)  # type: ignore[attr-defined]
+            for model_cls, col_name, _, is_json in results
+        }
+        assert ("internet_search_provider", "api_key", False) in found
+
+    def test_all_encrypted_string_columns_are_not_json(self) -> None:
+        results = _discover_encrypted_columns()
+        for model_cls, col_name, _, is_json in results:
+            col = getattr(model_cls, col_name).property.columns[0]
+            if isinstance(col.type, EncryptedString):
+                assert not is_json, (
+                    f"{model_cls.__tablename__}.{col_name} is EncryptedString "  # type: ignore[attr-defined]
+                    f"but is_json={is_json}"
+                )
+
+    def test_all_encrypted_json_columns_are_json(self) -> None:
+        results = _discover_encrypted_columns()
+        for model_cls, col_name, _, is_json in results:
+            col = getattr(model_cls, col_name).property.columns[0]
+            if isinstance(col.type, EncryptedJson):
+                assert is_json, (
+                    f"{model_cls.__tablename__}.{col_name} is EncryptedJson "  # type: ignore[attr-defined]
+                    f"but is_json={is_json}"
+                )
+
+
+class TestRotateCredential:
+    """Test rotation against the real Credential table (EncryptedJson).
+
+    Discovery is scoped to only the Credential model to avoid mutating
+    other tables in the test database.
+    """
+
+    @pytest.fixture(autouse=True)
+    def _limit_discovery(self) -> Generator[None, None, None]:
+        with patch(
+            f"{ROTATE_MODULE}._discover_encrypted_columns",
+            return_value=[(Credential, "credential_json", ["id"], True)],
+        ):
+            yield
+
+    @pytest.fixture()
+    def credential_id(
+        self, db_session: Session, tenant_context: None  # noqa: ARG002
+    ) -> Generator[int, None, None]:
+        """Insert a Credential row with raw encrypted bytes, clean up after."""
+        config = {"api_key": "sk-test-1234", "endpoint": "https://example.com"}
+        encrypted = _encrypt_string(json.dumps(config), key=OLD_KEY)
+
+        result = db_session.execute(
+            text(
+                "INSERT INTO credential "
+                "(source, credential_json, admin_public, curator_public) "
+                "VALUES (:source, :cred_json, true, false) "
+                "RETURNING id"
+            ),
+            {"source": DocumentSource.INGESTION_API.value, "cred_json": encrypted},
+        )
+        cred_id = result.scalar_one()
+        db_session.commit()
+
+        yield cred_id
+
+        db_session.execute(
+            text("DELETE FROM credential WHERE id = :id"), {"id": cred_id}
+        )
+        db_session.commit()
+
+    def test_rotates_credential_json(
+        self, db_session: Session, credential_id: int
+    ) -> None:
+        with (
+            patch(f"{ROTATE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+            patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+        ):
+            totals = rotate_encryption_key(db_session, old_key=OLD_KEY)
+
+        assert totals.get("credential.credential_json", 0) >= 1
+
+        raw = _raw_credential_bytes(db_session, credential_id)
+        assert raw is not None
+        decrypted = json.loads(_decrypt_bytes(raw, key=NEW_KEY))
+        assert decrypted["api_key"] == "sk-test-1234"
+        assert decrypted["endpoint"] == "https://example.com"
+
+    def test_skips_already_rotated(
+        self, db_session: Session, credential_id: int
+    ) -> None:
+        with (
+            patch(f"{ROTATE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+            patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+        ):
+            rotate_encryption_key(db_session, old_key=OLD_KEY)
+            _ = rotate_encryption_key(db_session, old_key=OLD_KEY)
+
+        raw = _raw_credential_bytes(db_session, credential_id)
+        assert raw is not None
+        decrypted = json.loads(_decrypt_bytes(raw, key=NEW_KEY))
+        assert decrypted["api_key"] == "sk-test-1234"
+
+    def test_dry_run_does_not_modify(
+        self, db_session: Session, credential_id: int
+    ) -> None:
+        original = _raw_credential_bytes(db_session, credential_id)
+
+        with (
+            patch(f"{ROTATE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+            patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+        ):
+            totals = rotate_encryption_key(db_session, old_key=OLD_KEY, dry_run=True)
+
+        assert totals.get("credential.credential_json", 0) >= 1
+
+        raw_after = _raw_credential_bytes(db_session, credential_id)
+        assert raw_after == original
+
+
+class TestRotateInternetSearchProvider:
+    """Test rotation against the real InternetSearchProvider table (EncryptedString).
+
+    Discovery is scoped to only the InternetSearchProvider model to avoid
+    mutating other tables in the test database.
+    """
+
+    @pytest.fixture(autouse=True)
+    def _limit_discovery(self) -> Generator[None, None, None]:
+        with patch(
+            f"{ROTATE_MODULE}._discover_encrypted_columns",
+            return_value=[
+                (InternetSearchProvider, "api_key", ["id"], False),
+            ],
+        ):
+            yield
+
+    @pytest.fixture()
+    def isp_id(
+        self, db_session: Session, tenant_context: None  # noqa: ARG002
+    ) -> Generator[int, None, None]:
+        """Insert an InternetSearchProvider row with raw encrypted bytes."""
+        encrypted = _encrypt_string("sk-secret-api-key", key=OLD_KEY)
+
+        result = db_session.execute(
+            text(
+                "INSERT INTO internet_search_provider "
+                "(name, provider_type, api_key, is_active) "
+                "VALUES (:name, :ptype, :api_key, false) "
+                "RETURNING id"
+            ),
+            {
+                "name": f"test-rotation-{id(self)}",
+                "ptype": "test",
+                "api_key": encrypted,
+            },
+        )
+        isp_id = result.scalar_one()
+        db_session.commit()
+
+        yield isp_id
+
+        db_session.execute(
+            text("DELETE FROM internet_search_provider WHERE id = :id"),
+            {"id": isp_id},
+        )
+        db_session.commit()
+
+    def test_rotates_api_key(self, db_session: Session, isp_id: int) -> None:
+        with (
+            patch(f"{ROTATE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+            patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+        ):
+            totals = rotate_encryption_key(db_session, old_key=OLD_KEY)
+
+        assert totals.get("internet_search_provider.api_key", 0) >= 1
+
+        raw = _raw_isp_bytes(db_session, isp_id)
+        assert raw is not None
+        assert _decrypt_bytes(raw, key=NEW_KEY) == "sk-secret-api-key"
+
+    def test_rotates_from_unencrypted(
+        self, db_session: Session, tenant_context: None  # noqa: ARG002
+    ) -> None:
+        """Test rotating data that was stored without any encryption key."""
+        result = db_session.execute(
+            text(
+                "INSERT INTO internet_search_provider "
+                "(name, provider_type, api_key, is_active) "
+                "VALUES (:name, :ptype, :api_key, false) "
+                "RETURNING id"
+            ),
+            {
+                "name": f"test-raw-{id(self)}",
+                "ptype": "test",
+                "api_key": b"raw-api-key",
+            },
+        )
+        isp_id = result.scalar_one()
+        db_session.commit()
+
+        try:
+            with (
+                patch(f"{ROTATE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+                patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", NEW_KEY),
+            ):
+                totals = rotate_encryption_key(db_session, old_key=None)
+
+            assert totals.get("internet_search_provider.api_key", 0) >= 1
+
+            raw = _raw_isp_bytes(db_session, isp_id)
+            assert raw is not None
+            assert _decrypt_bytes(raw, key=NEW_KEY) == "raw-api-key"
+        finally:
+            db_session.execute(
+                text("DELETE FROM internet_search_provider WHERE id = :id"),
+                {"id": isp_id},
+            )
+            db_session.commit()

backend/tests/unit/ee/conftest.py

@@ -0,0 +1,8 @@
+"""Auto-enable EE mode for all tests under tests/unit/ee/."""
+
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def _enable_ee_for_directory(enable_ee: None) -> None:  # noqa: ARG001
+    """Wraps the shared enable_ee fixture with autouse for this directory."""

backend/tests/unit/ee/onyx/utils/test_encryption.py

@@ -0,0 +1,165 @@
+"""Tests for EE AES-CBC encryption/decryption with explicit key support.
+
+With EE mode enabled (via conftest), fetch_versioned_implementation resolves
+to the EE implementations, so no patching of the MIT layer is needed.
+"""
+
+from unittest.mock import patch
+
+import pytest
+
+from ee.onyx.utils.encryption import _decrypt_bytes
+from ee.onyx.utils.encryption import _encrypt_string
+from ee.onyx.utils.encryption import _get_trimmed_key
+from ee.onyx.utils.encryption import decrypt_bytes_to_string
+from ee.onyx.utils.encryption import encrypt_string_to_bytes
+
+EE_MODULE = "ee.onyx.utils.encryption"
+
+# Keys must be exactly 16, 24, or 32 bytes for AES
+KEY_16 = "a" * 16
+KEY_16_ALT = "b" * 16
+KEY_24 = "d" * 24
+KEY_32 = "c" * 32
+
+
+@pytest.fixture(autouse=True)
+def _clear_key_cache() -> None:
+    _get_trimmed_key.cache_clear()
+
+
+class TestEncryptDecryptRoundTrip:
+    def test_roundtrip_with_env_key(self) -> None:
+        with patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", KEY_16):
+            encrypted = _encrypt_string("hello world")
+            assert encrypted != b"hello world"
+            assert _decrypt_bytes(encrypted) == "hello world"
+
+    def test_roundtrip_with_explicit_key(self) -> None:
+        encrypted = _encrypt_string("secret data", key=KEY_32)
+        assert encrypted != b"secret data"
+        assert _decrypt_bytes(encrypted, key=KEY_32) == "secret data"
+
+    def test_roundtrip_no_key(self) -> None:
+        """Without any key, data is raw-encoded (no encryption)."""
+        with patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", ""):
+            encrypted = _encrypt_string("plain text")
+            assert encrypted == b"plain text"
+            assert _decrypt_bytes(encrypted) == "plain text"
+
+    def test_explicit_key_overrides_env(self) -> None:
+        with patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", KEY_16):
+            encrypted = _encrypt_string("data", key=KEY_16_ALT)
+            with pytest.raises(ValueError):
+                _decrypt_bytes(encrypted, key=KEY_16)
+            assert _decrypt_bytes(encrypted, key=KEY_16_ALT) == "data"
+
+    def test_different_encryptions_produce_different_bytes(self) -> None:
+        """Each encryption uses a random IV, so results differ."""
+        a = _encrypt_string("same", key=KEY_16)
+        b = _encrypt_string("same", key=KEY_16)
+        assert a != b
+
+    def test_roundtrip_empty_string(self) -> None:
+        encrypted = _encrypt_string("", key=KEY_16)
+        assert encrypted != b""
+        assert _decrypt_bytes(encrypted, key=KEY_16) == ""
+
+    def test_roundtrip_unicode(self) -> None:
+        text = "日本語テスト 🔐 émojis"
+        encrypted = _encrypt_string(text, key=KEY_16)
+        assert _decrypt_bytes(encrypted, key=KEY_16) == text
+
+
+class TestDecryptFallbackBehavior:
+    def test_wrong_env_key_falls_back_to_raw_decode(self) -> None:
+        """Default key path: AES fails on non-AES data → fallback to raw decode."""
+        raw = "readable text".encode()
+        with patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", KEY_16):
+            assert _decrypt_bytes(raw) == "readable text"
+
+    def test_explicit_wrong_key_raises(self) -> None:
+        """Explicit key path: AES fails → raises, no fallback."""
+        encrypted = _encrypt_string("secret", key=KEY_16)
+        with pytest.raises(ValueError):
+            _decrypt_bytes(encrypted, key=KEY_16_ALT)
+
+    def test_explicit_none_key_with_no_env(self) -> None:
+        """key=None with empty env → raw decode."""
+        with patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", ""):
+            assert _decrypt_bytes(b"hello", key=None) == "hello"
+
+    def test_explicit_empty_string_key(self) -> None:
+        """key='' means no encryption."""
+        encrypted = _encrypt_string("test", key="")
+        assert encrypted == b"test"
+        assert _decrypt_bytes(encrypted, key="") == "test"
+
+
+class TestKeyValidation:
+    def test_key_too_short_raises(self) -> None:
+        with pytest.raises(RuntimeError, match="too short"):
+            _encrypt_string("data", key="short")
+
+    def test_16_byte_key(self) -> None:
+        encrypted = _encrypt_string("data", key=KEY_16)
+        assert _decrypt_bytes(encrypted, key=KEY_16) == "data"
+
+    def test_24_byte_key(self) -> None:
+        encrypted = _encrypt_string("data", key=KEY_24)
+        assert _decrypt_bytes(encrypted, key=KEY_24) == "data"
+
+    def test_32_byte_key(self) -> None:
+        encrypted = _encrypt_string("data", key=KEY_32)
+        assert _decrypt_bytes(encrypted, key=KEY_32) == "data"
+
+    def test_long_key_truncated_to_32(self) -> None:
+        """Keys longer than 32 bytes are truncated to 32."""
+        long_key = "e" * 64
+        encrypted = _encrypt_string("data", key=long_key)
+        assert _decrypt_bytes(encrypted, key=long_key) == "data"
+
+    def test_20_byte_key_trimmed_to_16(self) -> None:
+        """A 20-byte key is trimmed to the largest valid AES size that fits (16)."""
+        key_20 = "f" * 20
+        encrypted = _encrypt_string("data", key=key_20)
+        assert _decrypt_bytes(encrypted, key=key_20) == "data"
+
+        # Verify it was trimmed to 16 by checking that the first 16 bytes
+        # of the key can also decrypt it
+        key_16_same_prefix = "f" * 16
+        assert _decrypt_bytes(encrypted, key=key_16_same_prefix) == "data"
+
+    def test_25_byte_key_trimmed_to_24(self) -> None:
+        """A 25-byte key is trimmed to the largest valid AES size that fits (24)."""
+        key_25 = "g" * 25
+        encrypted = _encrypt_string("data", key=key_25)
+        assert _decrypt_bytes(encrypted, key=key_25) == "data"
+
+        key_24_same_prefix = "g" * 24
+        assert _decrypt_bytes(encrypted, key=key_24_same_prefix) == "data"
+
+    def test_30_byte_key_trimmed_to_24(self) -> None:
+        """A 30-byte key is trimmed to the largest valid AES size that fits (24)."""
+        key_30 = "h" * 30
+        encrypted = _encrypt_string("data", key=key_30)
+        assert _decrypt_bytes(encrypted, key=key_30) == "data"
+
+        key_24_same_prefix = "h" * 24
+        assert _decrypt_bytes(encrypted, key=key_24_same_prefix) == "data"
+
+
+class TestWrapperFunctions:
+    """Test encrypt_string_to_bytes / decrypt_bytes_to_string pass key through.
+
+    With EE mode enabled, the wrappers resolve to EE implementations automatically.
+    """
+
+    def test_wrapper_passes_key(self) -> None:
+        encrypted = encrypt_string_to_bytes("payload", key=KEY_16)
+        assert decrypt_bytes_to_string(encrypted, key=KEY_16) == "payload"
+
+    def test_wrapper_no_key_uses_env(self) -> None:
+        with patch(f"{EE_MODULE}.ENCRYPTION_KEY_SECRET", KEY_32):
+            encrypted = encrypt_string_to_bytes("payload")
+            assert decrypt_bytes_to_string(encrypted) == "payload"

backend/tests/unit/onyx/connectors/jira/test_jira_permission_sync.py

@@ -9,6 +9,8 @@ from onyx.connectors.jira.utils import JIRA_SERVER_API_VERSION
 from onyx.db.models import ConnectorCredentialPair
 from onyx.utils.sensitive import make_mock_sensitive_value
 
+pytestmark = pytest.mark.usefixtures("enable_ee")
+
 
 @pytest.fixture
 def mock_jira_cc_pair(

backend/tests/unit/onyx/utils/test_json_river.py

@@ -0,0 +1,394 @@
+"""Tests for the jsonriver incremental JSON parser."""
+
+import json
+
+import pytest
+
+from onyx.utils.jsonriver import JsonValue
+from onyx.utils.jsonriver import Parser
+
+
+def _all_deltas(chunks: list[str]) -> list[JsonValue]:
+    """Feed chunks one at a time and collect all emitted deltas."""
+    parser = Parser()
+    deltas: list[JsonValue] = []
+    for chunk in chunks:
+        deltas.extend(parser.feed(chunk))
+    deltas.extend(parser.finish())
+    return deltas
+
+
+class TestParseComplete:
+    """Parsing complete JSON in a single chunk."""
+
+    def test_simple_object(self) -> None:
+        deltas = _all_deltas(['{"a": 1}'])
+        assert any(r == {"a": 1.0} or r == {"a": 1} for r in deltas)
+
+    def test_simple_array(self) -> None:
+        deltas = _all_deltas(["[1, 2, 3]"])
+        assert any(isinstance(r, list) for r in deltas)
+
+    def test_simple_string(self) -> None:
+        deltas = _all_deltas(['"hello"'])
+        assert "hello" in deltas or any("hello" in str(r) for r in deltas)
+
+    def test_null(self) -> None:
+        deltas = _all_deltas(["null"])
+        assert None in deltas
+
+    def test_boolean_true(self) -> None:
+        deltas = _all_deltas(["true"])
+        assert True in deltas
+
+    def test_boolean_false(self) -> None:
+        deltas = _all_deltas(["false"])
+        assert any(r is False for r in deltas)
+
+    def test_number(self) -> None:
+        deltas = _all_deltas(["42"])
+        assert 42.0 in deltas
+
+    def test_negative_number(self) -> None:
+        deltas = _all_deltas(["-3.14"])
+        assert any(abs(r - (-3.14)) < 1e-10 for r in deltas if isinstance(r, float))
+
+    def test_empty_object(self) -> None:
+        deltas = _all_deltas(["{}"])
+        assert {} in deltas
+
+    def test_empty_array(self) -> None:
+        deltas = _all_deltas(["[]"])
+        assert [] in deltas
+
+
+class TestStreamingDeltas:
+    """Incremental feeding produces correct deltas."""
+
+    def test_object_string_value_streamed_char_by_char(self) -> None:
+        chunks = list('{"code": "abc"}')
+        deltas = _all_deltas(chunks)
+        str_parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "code" in d:
+                val = d["code"]
+                if isinstance(val, str):
+                    str_parts.append(val)
+        assert "".join(str_parts) == "abc"
+
+    def test_object_streamed_in_two_halves(self) -> None:
+        deltas = _all_deltas(['{"name": "Al', 'ice"}'])
+        str_parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "name" in d:
+                val = d["name"]
+                if isinstance(val, str):
+                    str_parts.append(val)
+        assert "".join(str_parts) == "Alice"
+
+    def test_multiple_keys_streamed(self) -> None:
+        deltas = _all_deltas(['{"a": "x', '", "b": "y"}'])
+        a_parts: list[str] = []
+        b_parts: list[str] = []
+        for d in deltas:
+            if isinstance(d, dict):
+                if "a" in d and isinstance(d["a"], str):
+                    a_parts.append(d["a"])
+                if "b" in d and isinstance(d["b"], str):
+                    b_parts.append(d["b"])
+        assert "".join(a_parts) == "x"
+        assert "".join(b_parts) == "y"
+
+    def test_deltas_only_contain_new_string_content(self) -> None:
+        parser = Parser()
+        d1 = parser.feed('{"msg": "hel')
+        d2 = parser.feed('lo"}')
+        parser.finish()
+
+        msg_parts = []
+        for d in d1 + d2:
+            if isinstance(d, dict) and "msg" in d:
+                val = d["msg"]
+                if isinstance(val, str):
+                    msg_parts.append(val)
+        assert "".join(msg_parts) == "hello"
+
+        # Each delta should only contain new chars, not repeat previous ones
+        if len(msg_parts) == 2:
+            assert msg_parts[0] == "hel"
+            assert msg_parts[1] == "lo"
+
+
+class TestEscapeSequences:
+    """JSON escape sequences are decoded correctly, even across chunk boundaries."""
+
+    def test_newline_escape(self) -> None:
+        deltas = _all_deltas(['{"text": "line1\\nline2"}'])
+        text_parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "text" in d and isinstance(d["text"], str):
+                text_parts.append(d["text"])
+        assert "".join(text_parts) == "line1\nline2"
+
+    def test_tab_escape(self) -> None:
+        deltas = _all_deltas(['{"t": "a\\tb"}'])
+        parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "t" in d and isinstance(d["t"], str):
+                parts.append(d["t"])
+        assert "".join(parts) == "a\tb"
+
+    def test_escaped_quote(self) -> None:
+        deltas = _all_deltas(['{"q": "say \\"hi\\""}'])
+        parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "q" in d and isinstance(d["q"], str):
+                parts.append(d["q"])
+        assert "".join(parts) == 'say "hi"'
+
+    def test_unicode_escape(self) -> None:
+        deltas = _all_deltas(['{"u": "\\u0041\\u0042"}'])
+        parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "u" in d and isinstance(d["u"], str):
+                parts.append(d["u"])
+        assert "".join(parts) == "AB"
+
+    def test_escape_split_across_chunks(self) -> None:
+        deltas = _all_deltas(['{"x": "a\\', 'nb"}'])
+        parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "x" in d and isinstance(d["x"], str):
+                parts.append(d["x"])
+        assert "".join(parts) == "a\nb"
+
+    def test_unicode_escape_split_across_chunks(self) -> None:
+        deltas = _all_deltas(['{"u": "\\u00', '41"}'])
+        parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "u" in d and isinstance(d["u"], str):
+                parts.append(d["u"])
+        assert "".join(parts) == "A"
+
+    def test_backslash_escape(self) -> None:
+        deltas = _all_deltas(['{"p": "c:\\\\dir"}'])
+        parts = []
+        for d in deltas:
+            if isinstance(d, dict) and "p" in d and isinstance(d["p"], str):
+                parts.append(d["p"])
+        assert "".join(parts) == "c:\\dir"
+
+
+class TestNestedStructures:
+    """Nested objects and arrays produce correct deltas."""
+
+    def test_nested_object(self) -> None:
+        deltas = _all_deltas(['{"outer": {"inner": "val"}}'])
+        found = False
+        for d in deltas:
+            if isinstance(d, dict) and "outer" in d:
+                outer = d["outer"]
+                if isinstance(outer, dict) and "inner" in outer:
+                    found = True
+        assert found
+
+    def test_array_of_strings(self) -> None:
+        deltas = _all_deltas(['["a', '", "b"]'])
+        all_items: list[str] = []
+        for d in deltas:
+            if isinstance(d, list):
+                for item in d:
+                    if isinstance(item, str):
+                        all_items.append(item)
+            elif isinstance(d, str):
+                all_items.append(d)
+        joined = "".join(all_items)
+        assert "a" in joined
+        assert "b" in joined
+
+    def test_object_with_number_and_bool(self) -> None:
+        deltas = _all_deltas(['{"count": 42, "active": true}'])
+        has_count = False
+        has_active = False
+        for d in deltas:
+            if isinstance(d, dict):
+                if "count" in d and d["count"] == 42.0:
+                    has_count = True
+                if "active" in d and d["active"] is True:
+                    has_active = True
+        assert has_count
+        assert has_active
+
+    def test_object_with_null_value(self) -> None:
+        deltas = _all_deltas(['{"key": null}'])
+        found = False
+        for d in deltas:
+            if isinstance(d, dict) and "key" in d and d["key"] is None:
+                found = True
+        assert found
+
+
+class TestComputeDelta:
+    """Direct tests for the _compute_delta static method."""
+
+    def test_none_prev_returns_current(self) -> None:
+        assert Parser._compute_delta(None, {"a": "b"}) == {"a": "b"}
+
+    def test_string_delta(self) -> None:
+        assert Parser._compute_delta("hel", "hello") == "lo"
+
+    def test_string_no_change(self) -> None:
+        assert Parser._compute_delta("same", "same") is None
+
+    def test_dict_new_key(self) -> None:
+        assert Parser._compute_delta({"a": "x"}, {"a": "x", "b": "y"}) == {"b": "y"}
+
+    def test_dict_string_append(self) -> None:
+        assert Parser._compute_delta({"code": "def"}, {"code": "def hello()"}) == {
+            "code": " hello()"
+        }
+
+    def test_dict_no_change(self) -> None:
+        assert Parser._compute_delta({"a": 1}, {"a": 1}) is None
+
+    def test_list_new_items(self) -> None:
+        assert Parser._compute_delta([1, 2], [1, 2, 3]) == [3]
+
+    def test_list_last_item_updated(self) -> None:
+        assert Parser._compute_delta(["a"], ["ab"]) == ["ab"]
+
+    def test_list_no_change(self) -> None:
+        assert Parser._compute_delta([1, 2], [1, 2]) is None
+
+    def test_primitive_change(self) -> None:
+        assert Parser._compute_delta(1, 2) == 2
+
+    def test_primitive_no_change(self) -> None:
+        assert Parser._compute_delta(42, 42) is None
+
+
+class TestParserLifecycle:
+    """Edge cases around parser state and lifecycle."""
+
+    def test_feed_after_finish_returns_empty(self) -> None:
+        parser = Parser()
+        parser.feed('{"a": 1}')
+        parser.finish()
+        assert parser.feed("more") == []
+
+    def test_empty_feed_returns_empty(self) -> None:
+        parser = Parser()
+        assert parser.feed("") == []
+
+    def test_whitespace_only_returns_empty(self) -> None:
+        parser = Parser()
+        assert parser.feed("   ") == []
+
+    def test_finish_with_trailing_whitespace(self) -> None:
+        parser = Parser()
+        # Trailing whitespace terminates the number, so feed() emits it
+        deltas = parser.feed("42  ")
+        assert 42.0 in deltas
+        parser.finish()  # Should not raise
+
+    def test_finish_with_trailing_content_raises(self) -> None:
+        parser = Parser()
+        # Feed a complete JSON value followed by non-whitespace in one chunk
+        parser.feed('{"a": 1} extra')
+        with pytest.raises(ValueError, match="Unexpected trailing"):
+            parser.finish()
+
+    def test_finish_flushes_pending_number(self) -> None:
+        parser = Parser()
+        deltas = parser.feed("42")
+        # Number has no terminator, so feed() can't emit it yet
+        assert deltas == []
+        final = parser.finish()
+        assert 42.0 in final
+
+
+class TestToolCallSimulation:
+    """Simulate the LLM tool-call streaming use case."""
+
+    def test_python_tool_call_streaming(self) -> None:
+        full_json = json.dumps({"code": "print('hello world')"})
+        chunk_size = 5
+        chunks = [
+            full_json[i : i + chunk_size] for i in range(0, len(full_json), chunk_size)
+        ]
+
+        parser = Parser()
+        code_parts: list[str] = []
+        for chunk in chunks:
+            for delta in parser.feed(chunk):
+                if isinstance(delta, dict) and "code" in delta:
+                    val = delta["code"]
+                    if isinstance(val, str):
+                        code_parts.append(val)
+        for delta in parser.finish():
+            if isinstance(delta, dict) and "code" in delta:
+                val = delta["code"]
+                if isinstance(val, str):
+                    code_parts.append(val)
+        assert "".join(code_parts) == "print('hello world')"
+
+    def test_multi_arg_tool_call(self) -> None:
+        full = '{"query": "search term", "num_results": 5}'
+        chunks = [full[:15], full[15:30], full[30:]]
+
+        parser = Parser()
+        query_parts: list[str] = []
+        has_num_results = False
+        for chunk in chunks:
+            for delta in parser.feed(chunk):
+                if isinstance(delta, dict):
+                    if "query" in delta and isinstance(delta["query"], str):
+                        query_parts.append(delta["query"])
+                    if "num_results" in delta:
+                        has_num_results = True
+        for delta in parser.finish():
+            if isinstance(delta, dict):
+                if "query" in delta and isinstance(delta["query"], str):
+                    query_parts.append(delta["query"])
+                if "num_results" in delta:
+                    has_num_results = True
+        assert "".join(query_parts) == "search term"
+        assert has_num_results
+
+    def test_code_with_newlines_and_escapes(self) -> None:
+        code = 'def greet(name):\n    print(f"Hello, {name}!")\n    return True'
+        full = json.dumps({"code": code})
+        chunk_size = 8
+        chunks = [full[i : i + chunk_size] for i in range(0, len(full), chunk_size)]
+
+        parser = Parser()
+        code_parts: list[str] = []
+        for chunk in chunks:
+            for delta in parser.feed(chunk):
+                if isinstance(delta, dict) and "code" in delta:
+                    val = delta["code"]
+                    if isinstance(val, str):
+                        code_parts.append(val)
+        for delta in parser.finish():
+            if isinstance(delta, dict) and "code" in delta:
+                val = delta["code"]
+                if isinstance(val, str):
+                    code_parts.append(val)
+        assert "".join(code_parts) == code
+
+    def test_single_char_streaming(self) -> None:
+        full = '{"key": "value"}'
+        parser = Parser()
+        key_parts: list[str] = []
+        for ch in full:
+            for delta in parser.feed(ch):
+                if isinstance(delta, dict) and "key" in delta:
+                    val = delta["key"]
+                    if isinstance(val, str):
+                        key_parts.append(val)
+        for delta in parser.finish():
+            if isinstance(delta, dict) and "key" in delta:
+                val = delta["key"]
+                if isinstance(val, str):
+                    key_parts.append(val)
+        assert "".join(key_parts) == "value"

backend/tests/unit/onyx/utils/test_vespa_query.py

@@ -20,8 +20,6 @@ from onyx.document_index.vespa_constants import TENANT_ID
 from onyx.document_index.vespa_constants import USER_PROJECT
 from shared_configs.configs import MULTI_TENANT
 
-# Import the function under test
-
 
 class TestBuildVespaFilters:
     def test_empty_filters(self) -> None:
@@ -179,11 +177,27 @@ class TestBuildVespaFilters:
         assert f"!({HIDDEN}=true) and " == result
 
     def test_user_project_filter(self) -> None:
-        """Test user project filtering (replacement for user folder IDs)."""
-        # Single project id
+        """Test user project filtering.
+
+        project_id alone does NOT trigger a knowledge scope restriction
+        (an agent with no explicit knowledge should search everything).
+        It only participates when explicit knowledge filters are present.
+        """
+        # project_id alone → no restriction
         filters = IndexFilters(access_control_list=[], project_id=789)
         result = build_vespa_filters(filters)
-        assert f'!({HIDDEN}=true) and ({USER_PROJECT} contains "789") and ' == result
+        assert f"!({HIDDEN}=true) and " == result
+
+        # project_id with user_file_ids → both OR'd
+        id1 = UUID("00000000-0000-0000-0000-000000000123")
+        filters = IndexFilters(
+            access_control_list=[], project_id=789, user_file_ids=[id1]
+        )
+        result = build_vespa_filters(filters)
+        assert (
+            f'!({HIDDEN}=true) and (({DOCUMENT_ID} contains "{str(id1)}") or ({USER_PROJECT} contains "789")) and '
+            == result
+        )
 
         # No project id
         filters = IndexFilters(access_control_list=[], project_id=None)
@@ -217,7 +231,11 @@ class TestBuildVespaFilters:
         )
 
     def test_combined_filters(self) -> None:
-        """Test combining multiple filter types."""
+        """Test combining multiple filter types.
+
+        Knowledge-scope filters (document_set, user_file_ids, project_id,
+        persona_id) are OR'd together, while all other filters are AND'd.
+        """
         id1 = UUID("00000000-0000-0000-0000-000000000123")
         filters = IndexFilters(
             access_control_list=["user1", "group1"],
@@ -231,7 +249,6 @@ class TestBuildVespaFilters:
 
         result = build_vespa_filters(filters)
 
-        # Build expected result piece by piece for readability
         expected = f"!({HIDDEN}=true) and "
         expected += (
             '(access_control_list contains "user1" or '
@@ -239,9 +256,13 @@ class TestBuildVespaFilters:
         )
         expected += f'({SOURCE_TYPE} contains "web") and '
         expected += f'({METADATA_LIST} contains "color{INDEX_SEPARATOR}red") and '
-        expected += f'({DOCUMENT_SETS} contains "set1") and '
-        expected += f'({DOCUMENT_ID} contains "{str(id1)}") and '
-        expected += f'({USER_PROJECT} contains "789") and '
+        # Knowledge scope filters are OR'd together
+        expected += (
+            f'(({DOCUMENT_SETS} contains "set1")'
+            f' or ({DOCUMENT_ID} contains "{str(id1)}")'
+            f' or ({USER_PROJECT} contains "789")'
+            f") and "
+        )
         cutoff_secs = int(datetime(2023, 1, 1, tzinfo=timezone.utc).timestamp())
         expected += f"!({DOC_UPDATED_AT} < {cutoff_secs}) and "
 
@@ -251,6 +272,32 @@ class TestBuildVespaFilters:
         result_no_trailing = build_vespa_filters(filters, remove_trailing_and=True)
         assert expected[:-5] == result_no_trailing  # Remove trailing " and "
 
+    def test_knowledge_scope_single_filter_not_wrapped(self) -> None:
+        """When only one knowledge-scope filter is present it should not
+        be wrapped in an extra OR group."""
+        filters = IndexFilters(access_control_list=[], document_set=["set1"])
+        result = build_vespa_filters(filters)
+        assert f'!({HIDDEN}=true) and ({DOCUMENT_SETS} contains "set1") and ' == result
+
+    def test_knowledge_scope_document_set_and_user_files_ored(self) -> None:
+        """Document set filter and user file IDs must be OR'd so that
+        connector documents (in the set) and user files (with specific
+        IDs) can both be found."""
+        id1 = UUID("00000000-0000-0000-0000-000000000123")
+        filters = IndexFilters(
+            access_control_list=[],
+            document_set=["engineering"],
+            user_file_ids=[id1],
+        )
+        result = build_vespa_filters(filters)
+        expected = (
+            f"!({HIDDEN}=true) and "
+            f'(({DOCUMENT_SETS} contains "engineering")'
+            f' or ({DOCUMENT_ID} contains "{str(id1)}")'
+            f") and "
+        )
+        assert expected == result
+
     def test_empty_or_none_values(self) -> None:
         """Test with empty or None values in filter lists."""
         # Empty strings in document set

cli/README.md

@@ -1,5 +1,8 @@
 # Onyx CLI
 
+[![Release CLI](https://github.com/onyx-dot-app/onyx/actions/workflows/release-cli.yml/badge.svg)](https://github.com/onyx-dot-app/onyx/actions/workflows/release-cli.yml)
+[![PyPI](https://img.shields.io/pypi/v/onyx-cli.svg)](https://pypi.org/project/onyx-cli/)
+
 A terminal interface for chatting with your [Onyx](https://github.com/onyx-dot-app/onyx) agent. Built with Go using [Bubble Tea](https://github.com/charmbracelet/bubbletea) for the TUI framework.
 
 ## Installation
@@ -28,7 +31,7 @@ Environment variables override config file values:
 
 | Variable | Required | Description |
 |----------|----------|-------------|
-| `ONYX_SERVER_URL` | No | Server base URL (default: `http://localhost:3000`) |
+| `ONYX_SERVER_URL` | No | Server base URL (default: `https://cloud.onyx.app`) |
 | `ONYX_API_KEY` | Yes | API key for authentication |
 | `ONYX_PERSONA_ID` | No | Default agent/persona ID |
 

tools/ods/hatch_build.py

@@ -26,20 +26,16 @@ class CustomBuildHook(BuildHookInterface):
 
         # Get config and environment
         binary_name = self.config["binary_name"]
-        tag = os.getenv("GITHUB_REF_NAME", "dev").removeprefix(f"{binary_name}/")
+        tag_prefix = self.config.get("tag_prefix", binary_name)
+        tag = os.getenv("GITHUB_REF_NAME", "dev").removeprefix(f"{tag_prefix}/")
         commit = os.getenv("GITHUB_SHA", "none")
 
         # Build the Go binary if it doesn't exist
         if not os.path.exists(binary_name):
             print(f"Building Go binary '{binary_name}'...")
+            ldflags = f"-X main.version={tag} -X main.commit={commit} -s -w"
             subprocess.check_call(  # noqa: S603
-                [
-                    "go",
-                    "build",
-                    f"-ldflags=-X main.version={tag} -X main.commit={commit} -s -w",
-                    "-o",
-                    binary_name,
-                ],
+                ["go", "build", f"-ldflags={ldflags}", "-o", binary_name],
             )
 
         build_data["shared_scripts"] = {binary_name: binary_name}

tools/ods/internal/_version.py

@@ -3,6 +3,9 @@ from __future__ import annotations
 import os
 import re
 
-_tag = os.environ.get("GITHUB_REF_NAME", "v0.0.0-dev").removeprefix("ods/")
+# Must match tag_prefix in pyproject.toml [tool.hatch.build.targets.wheel.hooks.custom]
+TAG_PREFIX: str = "ods"
+
+_tag = os.environ.get("GITHUB_REF_NAME", "v0.0.0-dev").removeprefix(f"{TAG_PREFIX}/")
 _match = re.search(r"v?(\d+\.\d+\.\d+)", _tag)
 __version__ = _match.group(1) if _match else "0.0.0"

tools/ods/pyproject.toml

@@ -14,7 +14,9 @@ keywords = [
 classifiers = [
     "Programming Language :: Go",
     "License :: OSI Approved :: MIT License",
-    "Operating System :: OS Independent",
+    "Operating System :: POSIX :: Linux",
+    "Operating System :: MacOS",
+    "Operating System :: Microsoft :: Windows",
 ]
 dynamic = ["version"]
 dependencies = [
@@ -27,7 +29,7 @@ dependencies = [
 Repository = "https://github.com/onyx-dot-app/onyx"
 
 [tool.hatch.build]
-include = ["go.mod", "go.sum", "main.go", "**/*.go", "**/*.py"]
+include = ["go.mod", "go.sum", "main.go", "**/*.go", "**/*.py", "README.md"]
 
 [tool.hatch.version]
 source = "code"
@@ -36,6 +38,7 @@ path = "internal/_version.py"
 [tool.hatch.build.targets.wheel.hooks.custom]
 path = "hatch_build.py"
 binary_name = "ods"
+tag_prefix = "ods"
 
 [tool.uv]
 managed = false

web/src/app/admin/api-key/OnyxApiKeyForm.tsx

@@ -6,7 +6,7 @@ import { Button } from "@opal/components";
 import { Disabled } from "@opal/core";
 import Text from "@/refresh-components/texts/Text";
 import InputTypeIn from "@/refresh-components/inputs/InputTypeIn";
-import InputComboBox from "@/refresh-components/inputs/InputComboBox";
+import InputSelect from "@/refresh-components/inputs/InputSelect";
 import { FormikField } from "@/refresh-components/form/FormikField";
 import { FormField } from "@/refresh-components/form/FormField";
 import { USER_ROLE_LABELS, UserRole } from "@/lib/types";
@@ -107,26 +107,25 @@ export default function OnyxApiKeyForm({
                     <FormField name="role" state={state} className="w-full">
                       <FormField.Label>Role:</FormField.Label>
                       <FormField.Control>
-                        <InputComboBox
+                        <InputSelect
                           value={field.value}
                           onValueChange={(value) => helper.setValue(value)}
-                          options={[
-                            {
-                              label: USER_ROLE_LABELS[UserRole.LIMITED],
-                              value: UserRole.LIMITED.toString(),
-                            },
-                            {
-                              label: USER_ROLE_LABELS[UserRole.BASIC],
-                              value: UserRole.BASIC.toString(),
-                            },
-                            {
-                              label: USER_ROLE_LABELS[UserRole.ADMIN],
-                              value: UserRole.ADMIN.toString(),
-                            },
-                          ]}
-                          placeholder="Select a role"
-                          strict
-                        />
+                        >
+                          <InputSelect.Trigger placeholder="Select a role" />
+                          <InputSelect.Content>
+                            <InputSelect.Item
+                              value={UserRole.LIMITED.toString()}
+                            >
+                              {USER_ROLE_LABELS[UserRole.LIMITED]}
+                            </InputSelect.Item>
+                            <InputSelect.Item value={UserRole.BASIC.toString()}>
+                              {USER_ROLE_LABELS[UserRole.BASIC]}
+                            </InputSelect.Item>
+                            <InputSelect.Item value={UserRole.ADMIN.toString()}>
+                              {USER_ROLE_LABELS[UserRole.ADMIN]}
+                            </InputSelect.Item>
+                          </InputSelect.Content>
+                        </InputSelect>
                       </FormField.Control>
                       <FormField.Description>
                         Select the role for this API key. Limited has access to

web/src/refresh-components/buttons/LineItem.tsx

@@ -57,6 +57,12 @@ export interface LineItemProps
     WithoutStyles<React.HTMLAttributes<HTMLDivElement>>,
     "children"
   > {
+  /**
+   * Whether the row should behave like a standalone interactive button.
+   * Set to false when nested inside another interactive primitive
+   * (e.g. Radix Select.Item) to avoid nested focus targets.
+   */
+  interactive?: boolean;
   // line-item variants
   strikethrough?: boolean;
   danger?: boolean;
@@ -131,6 +137,7 @@ export interface LineItemProps
  * - The component automatically adds a `data-selected="true"` attribute for custom styling
  */
 export default function LineItem({
+  interactive = true,
   selected,
   strikethrough,
   danger,
@@ -164,6 +171,11 @@ export default function LineItem({
   const emphasisKey = emphasized ? "emphasized" : "normal";
 
   const handleKeyDown = (e: React.KeyboardEvent<HTMLDivElement>) => {
+    if (!interactive) {
+      props.onKeyDown?.(e);
+      return;
+    }
+
     if (e.key === "Enter") {
       e.preventDefault();
       (e.currentTarget as HTMLDivElement).click();
@@ -174,6 +186,11 @@ export default function LineItem({
   };
 
   const handleKeyUp = (e: React.KeyboardEvent<HTMLDivElement>) => {
+    if (!interactive) {
+      props.onKeyUp?.(e);
+      return;
+    }
+
     if (e.key === " ") {
       e.preventDefault();
       (e.currentTarget as HTMLDivElement).click();
@@ -184,8 +201,8 @@ export default function LineItem({
   const content = (
     <div
       ref={ref}
-      role="button"
-      tabIndex={0}
+      role={interactive ? "button" : undefined}
+      tabIndex={interactive ? 0 : undefined}
       className={cn(
         "flex flex-row w-full items-start p-2 rounded-08 group/LineItem gap-2",
         !!(children && description) ? "items-start" : "items-center",

web/src/refresh-components/inputs/InputSelect.tsx

@@ -369,7 +369,7 @@ function InputSelectItem({
     <SelectPrimitive.Item
       ref={ref}
       value={value}
-      className="outline-none focus:outline-none"
+      className="outline-none focus:outline-none rounded-08 data-[highlighted]:bg-background-tint-02"
       onSelect={onClick}
     >
       {/* Hidden ItemText for Radix to track selection */}
@@ -383,7 +383,7 @@ function InputSelectItem({
         selected={isSelected}
         emphasized
         description={description}
-        onClick={noProp((event) => event.preventDefault())}
+        interactive={false}
       >
         {children}
       </LineItem>

web/src/refresh-pages/AppPage.tsx

@@ -193,7 +193,7 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
       onSubmit({
         message,
         currentMessageFiles,
-        deepResearch: deepResearchEnabled,
+        deepResearch: deepResearchEnabledForCurrentWorkflow,
       });
     }
   }
@@ -218,6 +218,8 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
     chatSessionId: currentChatSessionId,
     agentId: selectedAgent?.id,
   });
+  const deepResearchEnabledForCurrentWorkflow =
+    currentProjectId === null && deepResearchEnabled;
 
   const [presentingDocument, setPresentingDocument] =
     useState<MinimalOnyxDocument | null>(null);
@@ -435,10 +437,15 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
     onSubmit({
       message: lastUserMsg.message,
       currentMessageFiles: currentMessageFiles,
-      deepResearch: deepResearchEnabled,
+      deepResearch: deepResearchEnabledForCurrentWorkflow,
       messageIdToResend: lastUserMsg.messageId,
     });
-  }, [messageHistory, onSubmit, currentMessageFiles, deepResearchEnabled]);
+  }, [
+    messageHistory,
+    onSubmit,
+    currentMessageFiles,
+    deepResearchEnabledForCurrentWorkflow,
+  ]);
 
   const toggleDocumentSidebar = useCallback(() => {
     if (!documentSidebarVisible) {
@@ -458,7 +465,7 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
       onSubmit({
         message,
         currentMessageFiles,
-        deepResearch: deepResearchEnabled,
+        deepResearch: deepResearchEnabledForCurrentWorkflow,
       });
       if (showOnboarding || !onboardingDismissed) {
         finishOnboarding();
@@ -468,7 +475,7 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
       resetInputBar,
       onSubmit,
       currentMessageFiles,
-      deepResearchEnabled,
+      deepResearchEnabledForCurrentWorkflow,
       showOnboarding,
       onboardingDismissed,
       finishOnboarding,
@@ -503,7 +510,7 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
         onSubmit({
           message,
           currentMessageFiles,
-          deepResearch: deepResearchEnabled,
+          deepResearch: deepResearchEnabledForCurrentWorkflow,
         });
         if (showOnboarding || !onboardingDismissed) {
           finishOnboarding();
@@ -524,7 +531,7 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
       resetInputBar,
       onSubmit,
       currentMessageFiles,
-      deepResearchEnabled,
+      deepResearchEnabledForCurrentWorkflow,
       showOnboarding,
       onboardingDismissed,
       finishOnboarding,
@@ -709,7 +716,7 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
             >
               {/* Main content grid — 3 rows, animated */}
               <div
-                className="flex-1 w-full grid min-h-0 transition-[grid-template-rows] duration-150 ease-in-out"
+                className="flex-1 w-full grid min-h-0 px-4 transition-[grid-template-rows] duration-150 ease-in-out"
                 style={gridStyle}
               >
                 {/* ── Top row: ChatUI / WelcomeMessage / ProjectUI ── */}
@@ -732,7 +739,9 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
                       <ChatUI
                         liveAgent={liveAgent!}
                         llmManager={llmManager}
-                        deepResearchEnabled={deepResearchEnabled}
+                        deepResearchEnabled={
+                          deepResearchEnabledForCurrentWorkflow
+                        }
                         currentMessageFiles={currentMessageFiles}
                         setPresentingDocument={setPresentingDocument}
                         onSubmit={onSubmit}
@@ -828,7 +837,9 @@ export default function AppPage({ firstMessage }: ChatPageProps) {
                       />
                       <AppInputBar
                         ref={chatInputBarRef}
-                        deepResearchEnabled={deepResearchEnabled}
+                        deepResearchEnabled={
+                          deepResearchEnabledForCurrentWorkflow
+                        }
                         toggleDeepResearch={toggleDeepResearch}
                         filterManager={filterManager}
                         llmManager={llmManager}

web/src/sections/input/AppInputBar.tsx

@@ -148,7 +148,7 @@ const AppInputBar = React.memo(
       classification === "search";
 
     const { forcedToolIds, setForcedToolIds } = useForcedTools();
-    const { currentMessageFiles, setCurrentMessageFiles } =
+    const { currentMessageFiles, setCurrentMessageFiles, currentProjectId } =
       useProjectsContext();
 
     const currentIndexingFiles = useMemo(() => {
@@ -200,9 +200,17 @@ const AppInputBar = React.memo(
       const textarea = textAreaRef.current;
       if (!wrapper || !textarea) return;
 
+      // Reset so scrollHeight reflects actual content size
       wrapper.style.height = `${MIN_INPUT_HEIGHT}px`;
+
+      // scrollHeight doesn't include the wrapper's padding, so add it back
+      const wrapperStyle = getComputedStyle(wrapper);
+      const paddingTop = parseFloat(wrapperStyle.paddingTop);
+      const paddingBottom = parseFloat(wrapperStyle.paddingBottom);
+      const contentHeight = textarea.scrollHeight + paddingTop + paddingBottom;
+
       wrapper.style.height = `${Math.min(
-        Math.max(textarea.scrollHeight, MIN_INPUT_HEIGHT),
+        Math.max(contentHeight, MIN_INPUT_HEIGHT),
         MAX_INPUT_HEIGHT
       )}px`;
     }, [message, isSearchMode]);
@@ -358,13 +366,19 @@ const AppInputBar = React.memo(
     const showDeepResearch = useMemo(() => {
       const deepResearchGloballyEnabled =
         combinedSettings?.settings?.deep_research_enabled ?? true;
+      const isProjectWorkflow = currentProjectId !== null;
+
+      // TODO(@yuhong): Re-enable Deep Research in Projects workflow once it is fully supported.
+      // https://linear.app/onyx-app/issue/ENG-3818/re-enable-deep-research-in-projects
       return (
+        !isProjectWorkflow &&
         deepResearchGloballyEnabled &&
         hasSearchToolsAvailable(selectedAgent?.tools || [])
       );
     }, [
       selectedAgent?.tools,
       combinedSettings?.settings?.deep_research_enabled,
+      currentProjectId,
     ]);
 
     function handleKeyDownForPromptShortcuts(