DocChat AI 📄

A Retrieval-Augmented Generation (RAG) powered document Q&A application. Upload PDF or TXT documents and ask questions about them using semantic search + LLM reasoning.

The system retrieves relevant document chunks using FAISS vector search, generates grounded answers using the Groq LLM API, and maintains conversation memory across turns within a session.

---

🔗 Live Demo

👉 https://huggingface.co/spaces/GoldSharon/docchat-ai

⚠️ Hosted on Hugging Face Spaces (free CPU tier). The app may take 20–60 seconds to start if the Space is sleeping.

---

📸 Screenshots

1. Document Upload — Ready to Chat

2. RAG Response — Document Q&A

3. General Chat — No Document

---

🛠 Tech Stack

Layer	Technology
Backend	FastAPI (Python)
LLM	Groq API
Vector Database	FAISS
Embeddings	sentence-transformers
Memory	In-process session store
Frontend	HTML + CSS + Vanilla JS
Deployment	Hugging Face Spaces (Docker)

---

✨ Features

• Upload PDF or TXT documents
• Semantic search using FAISS vector embeddings
• Context-aware answers via RAG pipeline
• Conversation memory — the LLM remembers prior Q&A turns within a session
• Semantic summary intent detection — automatically fetches more chunks when you ask for an overview
• Strict grounded responses — answers are constrained to document content, no hallucination
• General chat mode when no document is selected
• Markdown formatted responses
• Automatic session reset on restart

---

🧠 How the RAG Pipeline Works

User uploads document
        ↓
Text extracted and split into chunks
        ↓
Chunks converted to embeddings (sentence-transformers)
        ↓
Stored in FAISS vector index
        ↓
User asks a question
        ↓
Semantic intent check (summary vs. factual)
        ↓
Question converted to embedding
        ↓
FAISS retrieves top-k relevant chunks
        ↓
Session memory (prior Q&A turns) retrieved
        ↓
Context + memory + question sent to Groq LLM
        ↓
LLM generates a grounded final answer
        ↓
Answer stored in session memory for future turns

---

💬 Conversation Memory

Each chat session is identified by a session_id. As you ask questions, the prior exchanges are stored in memory and injected into the LLM prompt on each subsequent turn. This enables follow-up questions like:

"Who is mentioned in section 2?" (next turn) "What did you say about that person?"

Memory is in-process and session-scoped — it resets when the server restarts.

---

🔍 Summary Intent Detection

When your question semantically matches phrases like:

• "summarize this document"
• "give me an overview"
• "what topics are covered"

...the system automatically retrieves 10 chunks with no relevance threshold, enabling a broad document summary rather than a narrow factual lookup.

---

🚀 Run Locally

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/docchat.git
cd docchat

2. Create virtual environment

python -m venv venv
source venv/bin/activate

Windows:

venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Set environment variables

Create .env:

GROQ_API_KEY=your_api_key_here
GROQ_MODEL=llama3-70b-8192
MIN_RELEVANCE_SCORE=1.0
CHUNK_SIZE=500
CHUNK_OVERLAP=50

Get a free key at 👉 https://console.groq.com

5. Run the application

uvicorn app.main:app --reload

6. Open the browser

http://localhost:8000

---

📁 Project Structure

docchat/
├── app/
│   ├── api/
│   │   ├── __init__.py
│   │   ├── models.py
│   │   ├── routes.py
│   │   └── upload_routes.py
│   ├── core/
│   │   ├── __init__.py
│   │   └── config.py
│   ├── services/
│   │   ├── __init__.py
│   │   ├── document_services.py
│   │   ├── faiss_services.py
│   │   ├── groq_service.py
│   │   ├── memory_service.py       ← NEW: session memory store
│   │   ├── ollama_service.py
│   │   └── rag_service.py          ← updated: memory + intent detection
│   ├── static/
│   │   ├── app.js
│   │   ├── index.html
│   │   └── style.css
│   └── main.py
├── Dockerfile
├── requirements.txt
└── README.md

---

🔑 Environment Variables

Variable	Description
GROQ_API_KEY	API key for Groq LLM
GROQ_MODEL	Model used for generation
MIN_RELEVANCE_SCORE	FAISS similarity distance threshold
CHUNK_SIZE	Document chunk size (characters)
CHUNK_OVERLAP	Overlap between chunks (characters)

---

📡 API Overview

Method	Endpoint	Description
POST	/chat	Ask a question (RAG or general)
POST	/upload	Upload a PDF or TXT document
GET	/health	Health check
GET	/stats	FAISS index statistics

Chat request body:

{
  "question": "What is the main topic of this document?",
  "document_id": "abc123",
  "session_id": "user-session-xyz"
}

---

☁️ Deployment

This project is deployed on Hugging Face Spaces using Docker.

Steps:

1. Create a Space and select Docker SDK
2. Add Dockerfile to the repo root
3. Push project via Git
4. Add secrets in Space settings:

GROQ_API_KEY
GROQ_MODEL

The Space automatically builds and deploys the application.

---

🤝 Contributing

1. Fork the repository
2. Create a feature branch
3. Commit your changes
4. Push to your fork
5. Open a Pull Request

---

📜 License

MIT License — free to use and modify.