RAG MongoDB + FAISS + PDF Chat

A small end‑to‑end Retrieval‑Augmented Generation (RAG) system:
upload a PDF, index it into MongoDB + FAISS, and chat with an LLM over its content through a web UI.

Features

• PDF upload (drag & drop or file picker)
• Text extraction + chunking
• Semantic search with FAISS
• Answer generation with GPT‑2
• MongoDB document storage
• FastAPI backend + simple HTML/JS frontend

Tech Stack

• Python, FastAPI, Uvicorn
• MongoDB
• FAISS (CPU)
• SentenceTransformers (all-MiniLM-L6-v2)
• Transformers (GPT‑2)
• PyPDF
• Vanilla HTML/CSS/JS

Project Structure

'''bash rag_mongo_faiss/ ├── app/ │ ├── main.py # FastAPI app (upload + ask endpoints, serve UI) │ ├── db/ │ │ └── mongo_client.py # MongoDB client │ ├── retrieval/ │ │ └── retriever.py # FAISS index + search │ ├── generation/ │ │ └── llm.py # GPT‑2 generation │ ├── utils/ │ │ └── pdf_parser.py # PDF parsing + chunking │ └── static/ │ └── index.html # Web UI (upload + chat) ├── requirements.txt ├── .gitignore └── README.md '''

Setup

1. Clone and create venv

git clone https://github.com/maryemchk/rag_mongo_faiss.git cd rag_mongo_faiss

python -m venv venv

Windows: venv\Scripts\activate

Linux/macOS: source venv/bin/activate

2. Install dependencies

pip install -r requirements.txt

3. Start MongoDB

Use a local MongoDB instance, e.g.:

mongod

(or a running MongoDB service).

The app uses:

• DB: rag_demo
• Collection: documents

(see app/db/mongo_client.py)

4. Run the app

python -m app.main

You should see Uvicorn running on http://0.0.0.0:8000.

Usage

1. Open your browser at:

   • http://localhost:8000

2. In the left panel:

   • Upload a PDF (drag & drop or click)
   • The backend:
     • extracts text with PyPDF
     • chunks it
     • stores chunks in MongoDB
     • rebuilds the FAISS index

3. In the chat box (right side):

   • Ask a question about the uploaded PDF
   • The backend:
     • embeds your question
     • retrieves similar chunks with FAISS
     • builds a context string
     • calls GPT‑2 to generate an answer

4. You can also inspect the API:

   • http://localhost:8000/docs

Key Components

Retrieval (app/retrieval/retriever.py)

• Encodes all document chunks with SentenceTransformers
• Builds an in‑memory FAISS index
• Supports rebuilding the index after new uploads
• search(query, k) returns top‑k similar chunks from MongoDB

Generation (app/generation/llm.py)

• Loads GPT‑2 via Hugging Face Transformers
• Builds a simple prompt:

Context:

Question: Answer:

• Generates a short answer from the context

PDF Handling (app/utils/pdf_parser.py)

• Extracts raw text per page with pypdf.PdfReader
• Splits text into overlapping chunks (size ~512 chars, small overlap)
• Returns a list of text chunks ready to embed and store

Endpoints

• GET /health – simple health check
• POST /upload – upload a PDF (multipart/form-data, field name file)
• POST /ask – ask a question ({"question": "...", "top_k": 3})
• GET /docs – FastAPI Swagger UI
• / – serves the HTML UI

Notes / Limitations

• GPT‑2 is small and can hallucinate; it’s used here as a lightweight local demo.
• FAISS index is rebuilt in memory when new PDFs are uploaded (okay for small demos).
• No auth, no multi‑user separation (demo only).

Possible Improvements

• Use a stronger instruction‑tuned model instead of GPT‑2
• Persist FAISS index to disk or use a vector DB
• Add document listing/deletion in the UI
• Better chunking (sentence/paragraph based)
• Hybrid retrieval (BM25 + FAISS) and reranking

---

This project is mainly for learning and portfolio: it shows end‑to‑end understanding of a basic RAG stack (ingestion → storage → retrieval → generation → UI).