📄 DocChat — RAG-based Document Chatbot

A Retrieval-Augmented Generation (RAG) chatbot that lets you upload any PDF and chat with it. Built with LangChain, FAISS, HuggingFace embeddings, Groq (Llama 3.3 70B), and Streamlit.

Problem Domain

Large Language Models are powerful but have two key limitations: they don't know your private/domain-specific documents, and they hallucinate when asked about content outside their training data. DocChat solves this by grounding every answer in the user's own documents — making it useful for querying research papers, company reports, manuals, legal documents, or study material without reading hundreds of pages manually.

Architecture

                 ┌─────────────────────────────────────────┐
                 │              INGESTION (once)           │
   PDF upload ──▶  PyPDFLoader ──▶ RecursiveCharacter     │
                 │                  TextSplitter (1000/200)│
                 │        │                                │
                 │        ▼                                │
                 │  HuggingFace Embeddings (MiniLM-L6-v2)  │
                 │        │                                │
                 │        ▼                                │
                 │   FAISS Vector Index (in-memory)        │
                 └─────────────────────────────────────────┘

                 ┌─────────────────────────────────────────┐
                 │            QUERY TIME (per turn)        │
   User query ──▶  Embed query ──▶ FAISS similarity       │
                 │                  search (top k=4)       │
                 │        │                                │
                 │        ▼                                │
                 │  Augmented prompt (context + history    │
                 │  + question)                            │
                 │        │                                │
                 │        ▼                                │
                 │  Groq LLM (Llama 3.3 70B) ──▶ Answer   │
                 └─────────────────────────────────────────┘

Design Choices

Component	Choice	Why
Orchestration	LangChain (LCEL)	Standard framework for composing RAG pipelines
LLM	Groq — Llama 3.3 70B	Free tier, extremely fast inference, strong open model
Embeddings	all-MiniLM-L6-v2 (local)	Free, fast, no API cost, good quality for semantic search
Vector store	FAISS	Lightweight, in-memory, no external DB needed
Chunking	1000 chars, 200 overlap	Balances context completeness vs retrieval precision
UI	Streamlit	Rapid chat UI with file upload and session state

Anti-hallucination: the system prompt restricts the model to answer only from retrieved context, and the UI shows the exact retrieved chunks (with file + page) under each answer for transparency.

How to Run

# 1. Clone and enter the project
git clone https://github.com/dishaa28/rag-chatbot.git
cd rag-chatbot

# 2. Create a virtual environment
python -m venv venv
venv\Scripts\activate        # Windows
# source venv/bin/activate   # Mac/Linux

# 3. Install dependencies
pip install -r requirements.txt

# 4. Add your Groq API key (free at console.groq.com)
copy .env.example .env       # then paste your key into .env
# (or just enter the key in the sidebar at runtime)

# 5. Launch
streamlit run app.py

Then open http://localhost:8501, upload one or more PDFs, click Process Documents, and start asking questions.

Project Structure

rag-chatbot/
├── app.py            # Streamlit chat UI
├── rag_pipeline.py   # Ingestion, chunking, embeddings, FAISS, RAG chain
├── requirements.txt
├── .env.example
└── README.md

Possible Extensions

• Persist the FAISS index to disk (vectorstore.save_local) to avoid re-indexing
• Support more formats (DOCX, TXT, web pages via WebBaseLoader)
• History-aware retrieval (rewrite follow-up questions before searching)
• Deploy on Streamlit Community Cloud / Hugging Face Spaces