[dvejsada]

Summary

Upstream: nothing to merge; upstream/main (LibreChat-AI/rag-api, latest 2026-09-23) is already an ancestor of main.

1. Dependency bumps (chore(deps))

• Every pin moved to its latest release within its current major line.
• The langchain family moves as one set (langchain-core 1.7.0, langchain 1.4.7, langchain-openai 1.7.2, langchain-aws 1.9.0, langchain-google-genai 4.4.2, langchain-huggingface 1.2.3, langchain-mongodb 0.12.1), because all of their new versions need core ≥ 1.6.
• Other notable bumps: fastapi 0.161.1, starlette 1.9.0, uvicorn 0.54.0, unstructured 0.27.34, pandas 2.3.3, sqlalchemy 2.0.55, pypdf 6.21.2, boto3 1.43.118, aiofiles 25.1.0 (needed by current unstructured-client).
• tenacity stays at 9.1.4, because google-genai requires <9.2.
• Held on purpose (breaking majors): pandas 3, SQLAlchemy 2.1, Pillow 13, chardet 7, sentence-transformers 6, mistralai 3, pgvector 0.5, asyncpg 0.32.

2. Removed dependencies the fork no longer uses

• rapidocr-onnxruntime, opencv-python-headless: nothing imports them, and OCR is remote via Mistral. Also removed the libgl1 and libglib2.0-0 apt packages from the full Dockerfile, which only they needed.
• networkx: nothing imports it; only an optional langchain-mongodb extra lists it.
• pypdf: removed from the runtime deps and moved to test_requirements.txt, since only a test fixture uses it. unstructured-client still pulls it in.
• mongomock: nothing references it.

3. OCR images embedded in PDFs (feat(ocr))

Mistral OCR returns only a placeholder (![img-0.jpeg](img-0.jpeg)) for a raster image embedded in a PDF page, such as a photo or scan of a page, and does not read the text inside it.

• The PDF request now also asks for the extracted images (include_image_base64, image_min_size, image_limit).
• Each image is OCR'd in its own call on a small thread pool.
• The recognized text replaces the placeholder on that image's page. Image ids are matched per page, and results are applied in page order.
• If one image's call fails, it logs a warning and the placeholder is kept. A failure of the main PDF call still raises.
• With the feature off, the request is exactly as before.
• New settings, documented in the README:
  • MISTRAL_OCR_EMBEDDED_IMAGES (default True)
  • MISTRAL_OCR_IMAGE_MIN_SIZE (100)
  • MISTRAL_OCR_MAX_EXTRACTED_IMAGES (200, minimum 1; how many images Mistral extracts, which bounds the response size)
  • MISTRAL_OCR_MAX_EMBEDDED_IMAGES (50, minimum 1; how many distinct images are OCR'd and billed, counted after the size filter and dedupe)
  • MISTRAL_OCR_IMAGE_MIN_PAGE_FRACTION (0.1; smaller images such as logos are skipped)
  • MISTRAL_OCR_IMAGE_CONCURRENCY (4; image OCR calls in flight across all documents and requests)
• Identical images (e.g. a logo on every page) are OCR'd only once per document.
• OCR calls retry on 429, 5xx and connection errors with backoff, using the SDK's RetryConfig.
• If a placeholder shares its line with other content (a table cell, list item or sentence), the text goes inline instead of as a new paragraph.
• ⚠️ Each distinct embedded image that is OCR'd is billed as one extra OCR page.

Test plan

• Lite deps plus test deps install on Python 3.12, and pip check is clean. A dry run of the full requirements.txt resolves.
• Unit tests in a Linux python:3.12 container: 453 passed, 1 skipped (the skip is expected).
• Integration tests (pgvector via testcontainers): 35 passed.
• New unit tests for the embedded-image path: placeholder replacement, request kwargs on and off, per-image failure isolation, cap, page-scoped ids, out-of-order completion, dict-shaped responses, dedupe, page-area filter, inline and table placement, and retries through the real SDK with a mocked HTTP transport.
• Not verified: a live Mistral OCR call on a real PDF with an embedded photo. The exact image_base64 format is handled both as a data URI and as bare base64. The page-area filter assumes the bbox and page dimensions use the same pixel space.
• CI is label-gated and has not run yet; add the CI label to run it.

Known, not addressed here

• The huggingfacetei embeddings provider imports langchain_huggingface, which the lite image doesn't install.

🤖 Generated with Claude Code