RAG Document Q&A Project

This repository implements an end-to-end Retrieval-Augmented Generation (RAG) workflow:

1. Load local documents.
2. Split them into chunks.
3. Convert chunks to embeddings.
4. Store vectors for similarity search.
5. Retrieve relevant context and generate answers with an LLM.

What Is Included

• Multi-format document loading from data/:
  • PDF (.pdf)
  • TXT (.txt)
  • CSV (.csv)
  • Excel (.xlsx)
  • Word (.docx)
  • JSON (.json)
• Chunking with RecursiveCharacterTextSplitter.
• Embedding generation with sentence-transformers (all-MiniLM-L6-v2).
• Vector search with FAISS (module pipeline under src/).
• LLM answer generation through Groq (langchain_groq).
• Notebook experiments including ChromaDB-based variants.

Repository Layout

RAG/
├─ data/
│  ├─ pdf/
│  ├─ text_files/
│  └─ vector_store/
├─ notebook/
│  ├─ document.ipynb
│  └─ pdf_loader.ipynb
├─ src/
│  ├─ __init__.py
│  ├─ data_loader.py
│  ├─ embedding.py
│  ├─ vectorstore.py
│  └─ search.py
├─ Document Q&A with RAG.ipynb
├─ requirements.txt
└─ README.md

Prerequisites

• Python 3.10 or newer
• Windows PowerShell (or equivalent shell)
• A Groq API key

Setup

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install --upgrade pip
pip install -r requirements.txt

Create a .env file in project root:

GROQ_API_KEY=your_groq_api_key

Quick Start

Step 1: Add your documents

Put files into data/ (for example data/pdf/ and data/text_files/).

Step 2: Build FAISS index

python src\vectorstore.py

Expected result:

• Documents are loaded and chunked.
• Embeddings are generated.
• Index and metadata are saved to faiss_store/.

Step 3: Ask a question

python src\search.py

Expected result:

• Existing FAISS index is loaded (or created if missing).
• Top relevant chunks are retrieved.
• Groq LLM returns a context-grounded summary.

Module Overview

• src/data_loader.py
  • Resolves data paths safely.
  • Loads supported document types into LangChain Document objects.
• src/embedding.py
  • Splits loaded docs into chunks.
  • Generates dense embeddings for each chunk.
• src/vectorstore.py
  • Builds, saves, loads, and queries a FAISS index.
• src/search.py
  • Connects vector retrieval with Groq LLM summarization.

Notebook Workflow

Use notebooks for experimentation and step-by-step debugging:

• notebook/document.ipynb for ingestion/loader checks.
• notebook/pdf_loader.ipynb for full PDF-to-RAG experimentation.

Note:

• src/ currently uses FAISS.
• Some notebook cells use ChromaDB.
• These are separate storage flows.

Troubleshooting

• GROQ_API_KEY not found
  • Confirm .env exists in project root and contains a valid key.
• Loader errors for specific formats
  • Verify optional parser dependencies are installed.
• Empty or weak retrieval results
  • Check documents actually loaded.
  • Increase top_k.
  • Lower score threshold.
  • Tune chunk size and overlap.

Suggested Next Improvements

• Add tests for loaders, chunking, and retrieval quality.
• Unify vector database choice across notebooks and modules.
• Add metadata-aware filtering (source, page, type).
• Add CLI or lightweight web app (for example, Streamlit) for interactive Q&A.