RAG + GraphQL Bonus Implementation (Replit-ready scaffold)

This repository is a scaffold for the bonus: a working RAG pipeline with re-ranking, retrieval-quality comparison, and basic latency/accuracy measurements.

It is designed to run on Replit / local machine. The code contains two modes:

• Provider mode: uses OpenAI for embeddings + LLM (requires OPENAI_API_KEY).
• Local mode: uses sentence-transformers (for embeddings) and cross-encoder for reranking. Local generation is left as a placeholder (you can plug a local LLM).

What is included

• rag_demo.py — main script: ingest, index (FAISS), search, rerank, prompt assembly, evaluate (LLM-only vs RAG).
• vector_store.py — helper: build FAISS index and search.
• reranker.py — helper: cross-encoder re-ranking wrapper.
• data/session_doc.txt — placeholder: drop your Session_3 text here (I could not auto-extract DOCX in this scaffold).
• requirements.txt — suggested packages.
• run_demo.sh — example run script.

Quickstart (Replit or local)

1. Create a repl / VM with internet access.
2. Upload your domain document to data/session_doc.txt (or other txt files in data/).
3. Option A (OpenAI):
   • Set OPENAI_API_KEY in repl secrets / environment.
   • pip install -r requirements.txt
   • python rag_demo.py --mode openai
4. Option B (Local embeddings & reranker):
   • pip install -r requirements.txt
   • python rag_demo.py --mode local
   • (This downloads sentence-transformers / cross-encoder models on first run.)

What the demo does

• Ingests text file(s) and chunks them (200–300 token chunks with overlap).
• Builds FAISS index from embeddings.
• Runs sample queries, measuring:
  • LLM-only (if configured) vs RAG answers (assembled prompt + LLM).
  • Retrieval latency, re-rank latency, LLM latency.
  • Simple retrieval quality metrics: provenance coverage (what percent of claimed lines include a substring from retrieved chunks).
• Outputs a comparison CSV in outputs/.

Notes

• This scaffold focuses on reproducible evaluation: you can replace placeholder LLM calls with any provider (OpenAI, local Llama, etc.).
• For production, swap FAISS for Pinecone/Weaviate and Apollo/Hasura for GraphQL layer.