melbinjp/DocQA

Cite the page in the answer, not only beside it

Refuse a bot check instead of storing it as the document

A document is a scan or it is not, and the reader gets a deadline

Give the manifest a budget, since my fixed number was wrong

Show enough of a document to say what it is, not who wrote it

Mark every chunk of a transcript, not just the first

Let the page reader fall back, and only spend a request where it is needed

Report why the page reader found nothing, instead of a bare 400

Look at the page, so a scan is a document and a chart is readable

Index a long chunk several times over, so a buried fact is reachable

Always say what is loaded, whatever retrieval returned

Tell the model which document each excerpt came from

Read HTML ourselves, instead of depending on a third party to do it

Index a table by its caption, and return its grid

Give a table its caption, or it cannot be found at all

Keep a table whole, and add a lexical retriever beside the dense one

Name the two retrieval caps, so raising one is not undone by the other

Stop cutting facts in half, and rank by cosine

Fetch the document itself, so a URL can be cited by page

Carry the page through extraction, so a citation can name one

Keep the free tiers awake, every three days

build: bump torch to 2.9.1; current transformers requires >=2.5 at import

build: resolve torch install deps from PyPI as fallback

fix: retry timeouts like high-demand errors, bound stream chunks, trim ingest 400 detail

Add MIT LICENSE

fix: Disable Accel-Buffering for SSE StreamingResponse

feat: Add SSRF protection and active Gemini fallback models

Increase embedding generation timeout to 180s

Integrate Jina Reader for URL content fetching with fallback

fix: Use BeautifulSoup directly for HTML/URL parsing to prevent timeouts on large pages

fix: Set standard browser User-Agent in backend HTTP client to prevent 403 Forbidden on URL ingestion

upg: Upgrade legacy google-generativeai to google-genai and migrate model to gemini-3.5-flash

fix: Implement retry mechanism with exponential backoff on LLM rate limits

fix: Robust parsing fallback for JSON/Form and PyMuPDF fallback for PDFs

fix: Add PDF/docx/pptx/xlsx extras to markitdown and support dynamic content-type parsing for ingest endpoint

fix: Correct PyTorch CPU installation index URL in Dockerfile

fix: Upgrade PyTorch to 2.4.0 in Dockerfile to support latest transformers and numpy 2.x

fix: Pin numpy<2.0.0 to prevent PyTorch 2.2.2 segfaults causing 503s

fix: Relax dependency version pins to resolve pip conflict with markitdown in Docker

feat: Replace PyPDF2 and custom parsers with MarkItDown for universal document ingestion

Merge pull request #6 from melbinjp/async-rag-improvements-1387425315086175663

feat: make RAG API fully non-blocking with proper error handling

Fix HF Hub download ReadTimeout error

Merge pull request #5 from melbinjp/feat/return-chunk-count

feat: Return chunk count in ingest response

Add session management endpoints for frontend integration

Update README.md

Merge pull request #4 from melbinjp/refine-and-enhance-app

Add Hugging Face Spaces configuration to README.md.

Merge pull request #3 from melbinjp/refine-and-enhance-app