RAGify Docs

RAGify Docs is a Retrieval-Augmented Generation prototype for loading local documents, building a FAISS vector index, and answering questions from retrieved context. The latest version also includes a frontend folder and a static GitHub Pages landing site in docs/.

Current Capabilities

• Loads PDFs, text files, CSVs, spreadsheets, Word documents, JSON, and markdown-ready document formats from the data/ directory.
• Chunks documents with LangChain's RecursiveCharacterTextSplitter.
• Embeds chunks with sentence-transformers/all-MiniLM-L6-v2.
• Persists and queries a local FAISS index.
• Uses Groq chat models for answer generation.
• Provides a FastAPI entrypoint in api.py.
• Includes a React frontend in frontend/.
• Includes a static GitHub Pages site in docs/.

Quickstart

Create a local .env file in the project root:

GROQ_API_KEY=your_groq_api_key_here

Run the CLI query path:

uv sync
uv run python app.py "What are hallucinations in LLMs?" --top-k 5

Rebuild the vector store before querying:

uv run python app.py "What is machine learning?" --rebuild

Run the API:

uv run uvicorn api:app --reload

Preview the static site locally:

python3 -m http.server 8765 --directory docs

Then open http://127.0.0.1:8765.

Project Layout

backend/
  data_loader.py    # Document ingestion
  embedding.py      # Chunking and embeddings
  vectorstore.py    # FAISS persistence and search
  search.py         # Retrieval plus LLM answering
frontend/           # React app
docs/               # Static GitHub Pages site
data/               # Local sample documents
api.py              # FastAPI endpoints
app.py              # CLI query runner

Recommended Improvements

• Normalize embeddings and use cosine similarity or inner-product FAISS search instead of raw L2 distance.
• Preserve source metadata during indexing, including file path, page number, and chunk id.
• Return citations with each generated answer so users can inspect the original evidence.
• Add an evaluation set with expected supporting documents and track recall@k, faithfulness, latency, and hallucination rate.
• Add hybrid retrieval with BM25 plus dense search, then rerank with a cross-encoder.
• Add tests for loaders, chunking, FAISS save/load, empty-index handling, API routes, and prompt construction.
• Keep generated artifacts out of git: .env, .DS_Store, __pycache__/, faiss_store/, and data/vector_store/.

GitHub Pages

The workflow in .github/workflows/pages.yml publishes the static site from docs/ whenever changes land on main. In the GitHub repository settings, set Pages to use GitHub Actions as its source.