DocChat — Chat with Your Documents 🤖📄

Upload any document. Ask anything. Get answers powered by 3 AI models simultaneously.

🔗 Live Demo: https://huggingface.co/spaces/loukikreddy22/docchat

---

🚀 What is DocChat?

DocChat is a full-stack RAG (Retrieval-Augmented Generation) application that lets you upload documents and have intelligent conversations with them. It uses 3 LLMs simultaneously and merges their answers for maximum accuracy.

---

✨ Features

• 📁 Upload PDF, DOCX, PPTX, TXT files
• 🧠 Multi-LLM fusion: Llama 3.3 + Gemini 2.0 Flash + Cohere Command-R
• 🌐 Web search augmentation via DuckDuckGo
• 🔍 Semantic search using FAISS vector store
• 💬 Session-based multi-file chat
• 🐳 Dockerized and deployed on HuggingFace Spaces

---

🏗️ System Architecture

User Uploads Document/File
(PDF / DOCX / PPTX / TXT)
                │
                ▼
      File Parsing Layer
      (Text Extraction)
                │
                ▼
      Text Chunking
      (500 Tokens + 50 Token Overlap)
                │
                ▼
      Embedding Model
      (Vector Generation)
                │
                ▼
      FAISS Vector Store
      (Semantic Indexing)
                │
                ▼
      User Query Input
                │
                ▼
      Semantic Retrieval
      (Top 4 Relevant Chunks)
                │
        ┌───────┴────────┐
        │                │
        ▼                ▼
Context Retrieval    Web Search
From Documents       External Sources
        │                │
        └───────┬────────┘
                │
                ▼
      Multi-LLM Layer
      • Llama
      • Gemini
      • Cohere
      (Parallel Inference)
                │
                ▼
      Response Fusion
      (Llama Aggregates Final Answer)
                │
                ▼
      Final AI Response
      (Returned to User)

⚙️ Workflow Summary

1. User uploads a document (PDF, DOCX, PPTX, or TXT)
2. The system extracts and preprocesses the document text
3. Text is split into overlapping chunks for better context retention
4. Embeddings are generated and stored in a FAISS vector database
5. User submits a question
6. Relevant chunks are retrieved using semantic similarity search
7. Additional context is fetched from:
   • Uploaded documents
   • External web sources
8. Multiple LLMs (Llama, Gemini, and Cohere) generate responses simultaneously
9. Llama combines and refines all generated outputs
10. Final intelligent response is returned to the user

---

🛠️ Tech Stack

Layer	Technology
Backend	FastAPI, Python
RAG Framework	LangChain
Vector Store	FAISS
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
LLMs	Groq/Llama 3.3, Gemini 2.0 Flash, Cohere Command-R
Web Search	DuckDuckGo Search
Frontend	HTML, CSS, Vanilla JS
Containerization	Docker
Deployment	HuggingFace Spaces

---

🔧 Run Locally

# Clone repo
git clone https://github.com/mloukikreddy/docchat
cd docchat

# Create virtual environment
python -m venv venv
venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

# Create .env file
GROQ_API_KEY=your_key
GEMINI_API_KEY=your_key
COHERE_API_KEY=your_key

# Run
uvicorn main:app --host 0.0.0.0 --port 8000

---

📁 Project Structure

docchat/
├── main.py           # FastAPI app, routes
├── rag_pipeline.py   # RAG logic, LLMs, FAISS
├── file_loader.py    # PDF/DOCX/PPTX/TXT parser
├── static/
│   └── index.html    # Frontend UI
├── Dockerfile        # Container config
├── requirements.txt
└── .env              # API keys (not committed)

---

🔑 Environment Variables

Variable	Description
GROQ_API_KEY	Get from console.groq.com
GEMINI_API_KEY	Get from aistudio.google.com
COHERE_API_KEY	Get from cohere.com

---

👤 Author

Mekala Loukik Reddy — AI & Data Science Student, Anurag University
LinkedIn | GitHub