README.md

metadata

title: Docs QA Backend
sdk: gradio
sdk_version: 5.49.1
python_version: '3.12'
app_file: app.py
pinned: false
app_port: 7860

RAG Backend (v0.3)

Setup

python -m venv venv
venv\Scripts\Activate.ps1          # Mac/Linux: source venv/bin/activate
pip install -r requirements.txt
copy .env.example .env             # add your free GROQ_API_KEY (https://console.groq.com/keys)
uvicorn app.main:app --reload

Docs UI: http://127.0.0.1:8000/docs

First start downloads two small models once (embedding ~90 MB, re-ranker ~90 MB). Data is stored in ./data. Upgrading from v0.2? Delete the old qdrant_data folder and upload your files again (chunking changed).

What it does

1. Upload (POST /documents) returns immediately; parsing, chunking and embedding run in the background with progress (GET /documents). Duplicate files are detected by hash. Pages with no text (scans) are reported.
2. Understands the question before searching: finds page numbers ("page 20", "20 no ke page"), document names, and rewrites the question into a clean English search query (Groq) so Hinglish and typos work.
3. Page questions ("what is on page 20?") read that page directly instead of guessing by similarity.
4. Other questions use hybrid search: vector search + BM25 keywords, fused with RRF, then a cross-encoder re-ranker. With several documents, each is searched separately so one large PDF cannot crowd out the others.
5. Answers come from Groq using only the retrieved context, with numbered citations.

Endpoints

• GET /health
• GET /documents, POST /documents, DELETE /documents/{doc_id}
• POST /ask with {"question": "...", "doc_ids": ["..."] | null, "history": [{"role": "user", "content": "..."}]}

Tests (no models or API key needed)

pip install pytest httpx reportlab
python -m pytest -q

Notes

• Page numbers are the position in the PDF file (1 = first page), which can differ from the number printed on the page.
• Scanned PDFs (images) have no text and need OCR first.
• Settings you can change in .env: LLM_MODEL, USE_RERANKER, USE_QUERY_REWRITE (see app/config.py for chunk size, top_k, etc.).