sonusaini209/crag_

Corrective RAG (CRAG)

Ask questions about a PDF/TXT file. Retrieved chunks are graded by an LLM; if the document can't answer the question, the system falls back to web search before answering. Based on Corrective Retrieval Augmented Generation.

How it works

question -> retrieve (FAISS, top 4) -> grade each chunk (0-1)
   CORRECT   (any chunk > 0.7)        -> refine -> generate
   AMBIGUOUS (mixed scores)           -> rewrite query -> Tavily web search -> refine (doc + web) -> generate
   INCORRECT (all chunks < 0.3)       -> rewrite query -> Tavily web search -> refine (web only)   -> generate

• Refine: context is split into sentences, de-duplicated, and each sentence is kept or dropped by an LLM filter.
• Stack: FastAPI, LangGraph, FAISS, MiniLM embeddings, Groq (openai/gpt-oss-120b), Tavily.
• The API returns the verdict, chunk scores, rewritten web query and sources, so you can see why CRAG chose its path.

Setup

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env        # add GROQ_API_KEY and TAVILY_API_KEY
uvicorn main:app --reload

Open http://127.0.0.1:8000, upload a file, ask a question.

API

• POST /upload (multipart file, .pdf or .txt, max 10 MB) -> {doc_id, chunks}
• POST /ask {question, doc_id} -> {answer, verdict, reason, scores, web_query, doc_sources, web_sources}
• GET /health

Limitations

• Indexes live in memory (last 20 documents), so they are lost on restart and don't work across multiple workers.
• Scanned PDFs (no text layer) are not supported.
• Single-turn Q&A; no chat history.