PDF RAG Chatbot

Live Demo: 👉 https://rag-chatbot-otcfoy7nvekby3tmmxycba.streamlit.app/

🎥 Demo

A conversational AI chatbot that lets you upload PDFs and ask questions about them. Built with Retrieval-Augmented Generation (RAG), it retrieves the most relevant chunks from your documents, reranks them for accuracy, and generates structured answers using LLaMA 3.3 70B via Groq.

---

Features

• Multi-PDF support — upload and index multiple PDFs in a single session
• Smart Mode — answers using document context, falls back to general knowledge when needed
• Strict Mode — answers strictly from the uploaded documents, no external knowledge
• Reranking pipeline — BGE cross-encoder reranker on top of MMR retrieval for higher precision
• Streaming responses — token-by-token output with a live typing cursor
• Query-type detection — automatically detects compare, summarize, definition, explain, or general queries and formats the answer accordingly
• Source attribution — shows which file and page each answer was pulled from, grouped by document
• Chat export — download the full conversation as a .txt file

---

How It Works

User Query
    |
Query Classifier --> Format Instructions
    |
History-Aware Query Rewriter  (LLaMA 3.3 via Groq)
    |
MMR Retriever  (ChromaDB + all-MiniLM-L6-v2)
fetch_k=40 --> k=20
    |
BGE Reranker  (BAAI/bge-reranker-base)
    |
Top 5 Chunks
    |
Answer Generation  (LLaMA 3.3 70B via Groq)
    |
Streaming Response + Source Expander

1. Ingest — PDFs are loaded with PyPDFLoader, split into 800-token chunks with 150-token overlap, embedded with all-MiniLM-L6-v2, and stored in ChromaDB.

2. Retrieve — On each query, chat history is used to rewrite the question into a clear standalone query. MMR retrieval fetches 40 candidates, filters to 20 for diversity, then the BGE cross-encoder reranks and returns the top 5.

3. Generate — The top chunks are passed to LLaMA 3.3 70B via Groq along with the detected query type and structured format instructions. The response streams token by token.

4. Sources — Retrieved chunks are grouped by source file and shown with page numbers and text excerpts in a collapsible expander.

---

Project Structure

rag-chatbot/
├── app.py              # Streamlit UI — chat interface, sidebar, export
├── rag_chain.py        # RAG pipeline — retriever, reranker, LLM chain
├── ingest.py           # PDF ingestion — load, chunk, embed, store
├── requirements.txt    # Python dependencies
├── .env                # API keys (not committed)
├── .gitignore
└── README.md

---

Getting Started

1. Clone the repo

git clone https://github.com/snigdha031/rag-chatbot.git
cd rag-chatbot

2. Create a virtual environment

python -m venv venv
source venv/bin/activate        # Mac/Linux
venv\Scripts\activate           # Windows

3. Install dependencies

pip install -r requirements.txt

4. Set up environment variables

Create a .env file in the root directory:

GROQ_API_KEY=your_groq_api_key_here

Get your free Groq API key at console.groq.com.

5. Run the app

streamlit run app.py

---

Configuration

Setting	Default	Description
Embedding model	all-MiniLM-L6-v2	HuggingFace sentence transformer
Reranker model	BAAI/bge-reranker-base	Cross-encoder reranker
LLM	llama-3.3-70b-versatile	Via Groq API
Chunk size	800 tokens	With 150 token overlap
MMR fetch_k	40	Candidates fetched before MMR
MMR k	20	After MMR diversity filter
Reranker top_k	5	Final chunks passed to LLM

---

Deployment

Streamlit Community Cloud (free)

1. Push the repo to GitHub
2. Go to share.streamlit.io
3. Connect your GitHub repo
4. Add your API key under Settings > Secrets:

GROQ_API_KEY = "your_groq_api_key_here"

5. Click Deploy — your app gets a public URL instantly

Note: ChromaDB persists to disk. On Streamlit Cloud the filesystem resets on each reboot, so uploaded PDFs will need to be re-indexed after the app restarts. For persistent storage across restarts, swap ChromaDB for a hosted vector DB like Pinecone or Qdrant Cloud.

---

Tech Stack

Layer	Tool
UI	Streamlit
LLM	LLaMA 3.3 70B via Groq
Orchestration	LangChain
Embeddings	all-MiniLM-L6-v2 (HuggingFace)
Vector store	ChromaDB
Reranker	BAAI/bge-reranker-base
PDF loading	PyPDFLoader

---

License

MIT License — see LICENSE for details.