📘 RAG Chat Assistant (Groq + ChromaDB)

A Retrieval-Augmented Generation (RAG) based document chat application built with Streamlit, Groq LLMs, and ChromaDB. This app allows users to upload documents (PDF, TXT, DOCX), index them efficiently, and ask natural language questions to get accurate, context-aware answers grounded in their own data.

---

RAG APP working Demo

Click here for the demo !

#1 (comment)

🚀 Features

• 📂 Upload and process multiple documents
• ✂️ Smart document chunking
• 🔎 Hybrid retrieval (BM25 + Vector Search)
• 🧠 Sentence-Transformers based embeddings
• ⚡ Groq-powered LLM inference (LLaMA models)
• 🗂 Persistent ChromaDB vector store
• 🧾 Source-aware answers
• 🎯 Optimized for enterprise document intelligence (e.g. lease analysis)

---

🏗️ Project Architecture

RAG App
│
├── app.py                  # Streamlit UI
├── modules/
│   ├── pipeline.py         # RAG orchestration logic
│   ├── loader.py           # File loading & chunking
│   ├── embedder.py         # Embedding manager
│   ├── retriever.py        # BM25 + vector retrieval
│   ├── llm.py              # Groq LLM wrapper
│   ├── config.py           # Central configuration
│                           # Helpers
│
├── chroma_db/              # Persistent vector store
├── requirements.txt
└── README.md

---

🔄 How the System Works (High Level)

graph TD
    A[User Uploads Documents] --> B[Document Loader]
    B --> C[Text Chunking]
    C --> D[Embeddings Generation]
    D --> E[ChromaDB Vector Store]

    F[User Query] --> G[Query Embedding]
    G --> H[Hybrid Retriever]
    E --> H
    H --> I[Relevant Chunks]
    I --> J[Groq LLM]
    J --> K[Final Answer]

Loading

---

🧠 Detailed RAG Pipeline Flow

sequenceDiagram
    participant U as User
    participant S as Streamlit App
    participant L as Loader
    participant E as Embedder
    participant C as ChromaDB
    participant R as Retriever
    participant G as Groq LLM

    U->>S: Upload Documents
    S->>L: Load & Parse Files
    L->>L: Chunk Text
    L->>E: Generate Embeddings
    E->>C: Store Vectors

    U->>S: Ask Question
    S->>E: Embed Query
    S->>R: Retrieve Relevant Chunks
    R->>C: Vector + BM25 Search
    R->>G: Context + Query
    G->>S: Grounded Answer
    S->>U: Display Response

Loading

---

📦 Installation

# Create virtual environment
python -m venv rven

# Activate (Windows)
rven\Scripts\activate

# Activate (Linux/Mac)
source rven/bin/activate

# Install dependencies
pip install -r requirements.txt

---

⚙️ Configuration

Update modules/config.py:

EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
CHUNK_SIZE = 800
CHUNK_OVERLAP = 100
GROQ_MODEL = "llama-3.1-8b-instant"

Set your Groq API key:

export GROQ_API_KEY=your_api_key_here

---

▶️ Run the Application

streamlit run app.py

---

🎯 Example Use Cases

• Lease agreement intelligence
• Policy & compliance document QA
• Internal knowledge base assistant
• Contract review automation
• Financial and legal document analysis

---

🛠️ Tech Stack

• Frontend: Streamlit
• LLM: Groq (LLaMA 3.1)
• Embeddings: Sentence-Transformers
• Vector DB: ChromaDB
• Search: BM25 + Dense Retrieval
• Language: Python 3.10+

---

🧩 Key Design Decisions

• Hybrid Retrieval improves recall and precision
• Chunk overlap prevents context loss
• Persistent vector store avoids re-indexing
• Model-agnostic design for easy upgrades

---

📌 Future Enhancements

• 🔐 User authentication
• 📊 Confidence scoring
• 🧠 Re-ranking with cross-encoders
• 📎 Highlight answers in source docs

---

🤝 Contributing

Pull requests are welcome. For major changes, please open an issue first.

👨‍💻 Author

Vikrant Singh Data Scientist | AI/ML Engineer Specialized in RAG Systems, NLP, and Enterprise AI

---

⭐ If this project helped you, consider starring the repository!