fix(image): Cap Uploaded File Image Count (#10298)

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

backend/onyx/configs/app_configs.py

@@ -788,6 +788,29 @@ MAX_FILE_SIZE_BYTES = int(
     os.environ.get("MAX_FILE_SIZE_BYTES") or 2 * 1024 * 1024 * 1024
 )  # 2GB in bytes
 
+# Maximum embedded images allowed in a single file. PDFs (and other formats)
+# with thousands of embedded images can OOM the user-file-processing worker
+# because every image is decoded with PIL and then sent to the vision LLM.
+# Enforced both at upload time (rejects the file) and during extraction
+# (defense-in-depth: caps the number of images materialized).
+#
+# Clamped to >= 0; a negative env value would turn upload validation into
+# always-fail and extraction into always-stop, which is never desired. 0
+# disables image extraction entirely, which is a valid (if aggressive) setting.
+MAX_EMBEDDED_IMAGES_PER_FILE = max(
+    0, int(os.environ.get("MAX_EMBEDDED_IMAGES_PER_FILE") or 500)
+)
+
+# Maximum embedded images allowed across all files in a single upload batch.
+# Protects against the scenario where a user uploads many files that each
+# fall under MAX_EMBEDDED_IMAGES_PER_FILE but aggregate to enough work
+# (serial-ish celery fan-out plus per-image vision-LLM calls) to OOM the
+# worker under concurrency or run up surprise latency/cost. Also clamped
+# to >= 0.
+MAX_EMBEDDED_IMAGES_PER_UPLOAD = max(
+    0, int(os.environ.get("MAX_EMBEDDED_IMAGES_PER_UPLOAD") or 1000)
+)
+
 # Use document summary for contextual rag
 USE_DOCUMENT_SUMMARY = os.environ.get("USE_DOCUMENT_SUMMARY", "true").lower() == "true"
 # Use chunk summary for contextual rag

backend/onyx/file_processing/extract_file_text.py

@@ -21,6 +21,7 @@ import chardet
 import openpyxl
 from PIL import Image
 
+from onyx.configs.app_configs import MAX_EMBEDDED_IMAGES_PER_FILE
 from onyx.configs.constants import ONYX_METADATA_FILENAME
 from onyx.configs.llm_configs import get_image_extraction_and_analysis_enabled
 from onyx.file_processing.file_types import OnyxFileExtensions
@@ -176,6 +177,56 @@ def read_text_file(
     return file_content_raw, metadata
 
 
+def count_pdf_embedded_images(file: IO[Any], cap: int) -> int:
+    """Return the number of embedded images in a PDF, short-circuiting at cap+1.
+
+    Used to reject PDFs whose image count would OOM the user-file-processing
+    worker during indexing. Returns a value > cap as a sentinel once the count
+    exceeds the cap, so callers do not iterate thousands of image objects just
+    to report a number. Returns 0 if the PDF cannot be parsed.
+
+    Owner-password-only PDFs (permission restrictions but no open password) are
+    counted normally — they decrypt with an empty string. Truly password-locked
+    PDFs are skipped (return 0) since we can't inspect them; the caller should
+    ensure the password-protected check runs first.
+
+    Always restores the file pointer to its original position before returning.
+    """
+    from pypdf import PdfReader
+
+    try:
+        start_pos = file.tell()
+    except Exception:
+        start_pos = None
+    try:
+        if start_pos is not None:
+            file.seek(0)
+        reader = PdfReader(file)
+        if reader.is_encrypted:
+            # Try empty password first (owner-password-only PDFs); give up if that fails.
+            try:
+                if reader.decrypt("") == 0:
+                    return 0
+            except Exception:
+                return 0
+        count = 0
+        for page in reader.pages:
+            for _ in page.images:
+                count += 1
+                if count > cap:
+                    return count
+        return count
+    except Exception:
+        logger.warning("Failed to count embedded images in PDF", exc_info=True)
+        return 0
+    finally:
+        if start_pos is not None:
+            try:
+                file.seek(start_pos)
+            except Exception:
+                pass
+
+
 def pdf_to_text(file: IO[Any], pdf_pass: str | None = None) -> str:
     """
     Extract text from a PDF. For embedded images, a more complex approach is needed.
@@ -231,8 +282,27 @@ def read_pdf_file(
         )
 
         if extract_images:
+            image_cap = MAX_EMBEDDED_IMAGES_PER_FILE
+            images_processed = 0
+            cap_reached = False
             for page_num, page in enumerate(pdf_reader.pages):
+                if cap_reached:
+                    break
                 for image_file_object in page.images:
+                    if images_processed >= image_cap:
+                        # Defense-in-depth backstop. Upload-time validation
+                        # should have rejected files exceeding the cap, but
+                        # we also break here so a single oversized file can
+                        # never pin a worker.
+                        logger.warning(
+                            "PDF embedded image cap reached (%d). "
+                            "Skipping remaining images on page %d and beyond.",
+                            image_cap,
+                            page_num + 1,
+                        )
+                        cap_reached = True
+                        break
+
                     image = Image.open(io.BytesIO(image_file_object.data))
                     img_byte_arr = io.BytesIO()
                     image.save(img_byte_arr, format=image.format)
@@ -245,6 +315,7 @@ def read_pdf_file(
                         image_callback(img_bytes, image_name)
                     else:
                         extracted_images.append((img_bytes, image_name))
+                    images_processed += 1
 
         return text, metadata, extracted_images
 

backend/onyx/server/features/build/api/user_library.py

@@ -40,6 +40,8 @@ from sqlalchemy.orm import Session
 
 from onyx.auth.users import current_user
 from onyx.background.celery.versioned_apps.client import app as celery_app
+from onyx.configs.app_configs import MAX_EMBEDDED_IMAGES_PER_FILE
+from onyx.configs.app_configs import MAX_EMBEDDED_IMAGES_PER_UPLOAD
 from onyx.configs.constants import DocumentSource
 from onyx.configs.constants import OnyxCeleryQueues
 from onyx.configs.constants import OnyxCeleryTask
@@ -50,6 +52,9 @@ from onyx.db.engine.sql_engine import get_session
 from onyx.db.enums import ConnectorCredentialPairStatus
 from onyx.db.models import User
 from onyx.document_index.interfaces import DocumentMetadata
+from onyx.error_handling.error_codes import OnyxErrorCode
+from onyx.error_handling.exceptions import OnyxError
+from onyx.file_processing.extract_file_text import count_pdf_embedded_images
 from onyx.server.features.build.configs import USER_LIBRARY_MAX_FILE_SIZE_BYTES
 from onyx.server.features.build.configs import USER_LIBRARY_MAX_FILES_PER_UPLOAD
 from onyx.server.features.build.configs import USER_LIBRARY_MAX_TOTAL_SIZE_BYTES
@@ -127,6 +132,49 @@ class DeleteFileResponse(BaseModel):
 # =============================================================================
 
 
+def _looks_like_pdf(filename: str, content_type: str | None) -> bool:
+    """True if either the filename or the content-type indicates a PDF.
+
+    Client-supplied ``content_type`` can be spoofed (e.g. a PDF uploaded with
+    ``Content-Type: application/octet-stream``), so we also fall back to
+    extension-based detection via ``mimetypes.guess_type`` on the filename.
+    """
+    if content_type == "application/pdf":
+        return True
+    guessed, _ = mimetypes.guess_type(filename)
+    return guessed == "application/pdf"
+
+
+def _check_pdf_image_caps(
+    filename: str, content: bytes, content_type: str | None, batch_total: int
+) -> int:
+    """Enforce per-file and per-batch embedded-image caps for PDFs.
+
+    Returns the number of embedded images in this file (0 for non-PDFs) so
+    callers can update their running batch total. Raises OnyxError(INVALID_INPUT)
+    if either cap is exceeded.
+    """
+    if not _looks_like_pdf(filename, content_type):
+        return 0
+    file_cap = MAX_EMBEDDED_IMAGES_PER_FILE
+    batch_cap = MAX_EMBEDDED_IMAGES_PER_UPLOAD
+    # Short-circuit at the larger cap so we get a useful count for both checks.
+    count = count_pdf_embedded_images(BytesIO(content), max(file_cap, batch_cap))
+    if count > file_cap:
+        raise OnyxError(
+            OnyxErrorCode.INVALID_INPUT,
+            f"PDF '{filename}' contains too many embedded images "
+            f"(more than {file_cap}). Try splitting the document into smaller files.",
+        )
+    if batch_total + count > batch_cap:
+        raise OnyxError(
+            OnyxErrorCode.INVALID_INPUT,
+            f"Upload would exceed the {batch_cap}-image limit across all "
+            f"files in this batch. Try uploading fewer image-heavy files at once.",
+        )
+    return count
+
+
 def _sanitize_path(path: str) -> str:
     """Sanitize a file path, removing traversal attempts and normalizing.
 
@@ -356,6 +404,7 @@ async def upload_files(
 
     uploaded_entries: list[LibraryEntryResponse] = []
     total_size = 0
+    batch_image_total = 0
     now = datetime.now(timezone.utc)
 
     # Sanitize the base path
@@ -375,6 +424,14 @@ async def upload_files(
                 detail=f"File '{file.filename}' exceeds maximum size of {USER_LIBRARY_MAX_FILE_SIZE_BYTES // (1024*1024)}MB",
             )
 
+        # Reject PDFs with an unreasonable per-file or per-batch image count
+        batch_image_total += _check_pdf_image_caps(
+            filename=file.filename or "unnamed",
+            content=content,
+            content_type=file.content_type,
+            batch_total=batch_image_total,
+        )
+
         # Validate cumulative storage (existing + this upload batch)
         total_size += file_size
         if existing_usage + total_size > USER_LIBRARY_MAX_TOTAL_SIZE_BYTES:
@@ -473,6 +530,7 @@ async def upload_zip(
 
     uploaded_entries: list[LibraryEntryResponse] = []
     total_size = 0
+    batch_image_total = 0
 
     # Extract zip contents into a subfolder named after the zip file
     zip_name = api_sanitize_filename(file.filename or "upload")
@@ -511,6 +569,36 @@ async def upload_zip(
                     logger.warning(f"Skipping '{zip_info.filename}' - exceeds max size")
                     continue
 
+                # Skip PDFs that would trip the per-file or per-batch image
+                # cap (would OOM the user-file-processing worker). Matches
+                # /upload behavior but uses skip-and-warn to stay consistent
+                # with the zip path's handling of oversized files.
+                zip_file_name = zip_info.filename.split("/")[-1]
+                zip_content_type, _ = mimetypes.guess_type(zip_file_name)
+                if zip_content_type == "application/pdf":
+                    image_count = count_pdf_embedded_images(
+                        BytesIO(file_content),
+                        max(
+                            MAX_EMBEDDED_IMAGES_PER_FILE,
+                            MAX_EMBEDDED_IMAGES_PER_UPLOAD,
+                        ),
+                    )
+                    if image_count > MAX_EMBEDDED_IMAGES_PER_FILE:
+                        logger.warning(
+                            "Skipping '%s' - exceeds %d per-file embedded-image cap",
+                            zip_info.filename,
+                            MAX_EMBEDDED_IMAGES_PER_FILE,
+                        )
+                        continue
+                    if batch_image_total + image_count > MAX_EMBEDDED_IMAGES_PER_UPLOAD:
+                        logger.warning(
+                            "Skipping '%s' - would exceed %d per-batch embedded-image cap",
+                            zip_info.filename,
+                            MAX_EMBEDDED_IMAGES_PER_UPLOAD,
+                        )
+                        continue
+                    batch_image_total += image_count
+
                 total_size += file_size
 
                 # Validate cumulative storage

backend/onyx/server/features/projects/projects_file_utils.py

@@ -10,7 +10,10 @@ from pydantic import Field
 from sqlalchemy.orm import Session
 
 from onyx.configs.app_configs import FILE_TOKEN_COUNT_THRESHOLD
+from onyx.configs.app_configs import MAX_EMBEDDED_IMAGES_PER_FILE
+from onyx.configs.app_configs import MAX_EMBEDDED_IMAGES_PER_UPLOAD
 from onyx.db.llm import fetch_default_llm_model
+from onyx.file_processing.extract_file_text import count_pdf_embedded_images
 from onyx.file_processing.extract_file_text import extract_file_text
 from onyx.file_processing.extract_file_text import get_file_ext
 from onyx.file_processing.file_types import OnyxFileExtensions
@@ -156,6 +159,11 @@ def categorize_uploaded_files(
         except RuntimeError as e:
             logger.warning(f"Failed to get current tenant ID: {str(e)}")
 
+    # Running total of embedded images across PDFs in this batch. Once the
+    # aggregate cap is reached, subsequent PDFs in the same upload are
+    # rejected even if they'd individually fit under MAX_EMBEDDED_IMAGES_PER_FILE.
+    batch_image_total = 0
+
     for upload in files:
         try:
             filename = get_safe_filename(upload)
@@ -204,6 +212,47 @@ def categorize_uploaded_files(
                     )
                     continue
 
+                # Reject PDFs with an unreasonable number of embedded images
+                # (either per-file or accumulated across this upload batch).
+                # A PDF with thousands of embedded images can OOM the
+                # user-file-processing celery worker because every image is
+                # decoded with PIL and then sent to the vision LLM.
+                if extension == ".pdf":
+                    file_cap = MAX_EMBEDDED_IMAGES_PER_FILE
+                    batch_cap = MAX_EMBEDDED_IMAGES_PER_UPLOAD
+                    # Use the larger of the two caps as the short-circuit
+                    # threshold so we get a useful count for both checks.
+                    # count_pdf_embedded_images restores the stream position.
+                    count = count_pdf_embedded_images(
+                        upload.file, max(file_cap, batch_cap)
+                    )
+                    if count > file_cap:
+                        results.rejected.append(
+                            RejectedFile(
+                                filename=filename,
+                                reason=(
+                                    f"PDF contains too many embedded images "
+                                    f"(more than {file_cap}). Try splitting "
+                                    f"the document into smaller files."
+                                ),
+                            )
+                        )
+                        continue
+                    if batch_image_total + count > batch_cap:
+                        results.rejected.append(
+                            RejectedFile(
+                                filename=filename,
+                                reason=(
+                                    f"Upload would exceed the "
+                                    f"{batch_cap}-image limit across all "
+                                    f"files in this batch. Try uploading "
+                                    f"fewer image-heavy files at once."
+                                ),
+                            )
+                        )
+                        continue
+                    batch_image_total += count
+
                 text_content = extract_file_text(
                     file=upload.file,
                     file_name=filename,

web/src/app/craft/v1/configure/components/UserLibraryModal.tsx

@@ -272,6 +272,22 @@ export default function UserLibraryModal({
                   </Disabled>
                 </Section>
 
+                {/* The exact cap is controlled by the backend env var
+                    MAX_EMBEDDED_IMAGES_PER_FILE (default 500). This copy is
+                    deliberately vague so it doesn't drift if the limit is
+                    tuned per-deployment; the precise number is surfaced in
+                    the rejection error the server returns. */}
+                <Section
+                  flexDirection="row"
+                  justifyContent="end"
+                  padding={0.5}
+                  height="fit"
+                >
+                  <Text secondaryBody text03>
+                    PDFs with many embedded images may be rejected.
+                  </Text>
+                </Section>
+
                 {isLoading ? (
                   <Section padding={2} height="fit">
                     <Text secondaryBody text03>