📄 Multimodal RAG Chatbot (Groq + MongoDB Atlas)

A Retrieval-Augmented Generation (RAG) application for asking questions about PDF documents. During ingestion, it extracts PDF text and uses a Groq-hosted vision model to turn charts and images into searchable text descriptions.

🚀 Key Features

• Semantic Text Search: Uses local all-MiniLM-L6-v2 embeddings for semantic retrieval.
• Multimodal Ingestion: Extracts PDF images and captions them with a configurable Groq vision model.
• Vector Database: Stores text chunks, image descriptions, embeddings, and metadata in MongoDB Atlas Vector Search.
• Configurable Generation: Uses a configurable Groq-hosted model to answer questions from retrieved context.
• Source Citations: Appends deduplicated PDF filenames and page numbers from retrieved metadata.
• Streamlit UI: A clean, modern chat interface for easy document management and interaction.

🛠️ Tech Stack

• AI Models: Groq (qwen/qwen3.6-27b by default), Hugging Face (all-MiniLM-L6-v2 local embeddings).
• Database: MongoDB Atlas (Vector Search).
• Orchestration: LangChain.
• Interface: Streamlit.
• PDF Processing: PyMuPDF (fitz), PyPDF.

📋 Architecture

1. Ingestion: PDFs are split into text chunks. Images are extracted and sent to Groq Vision for detailed captioning.
2. Indexing: Both text and image captions are converted into 384-dimensional vectors and stored in MongoDB Atlas.
3. Retrieval: When a user asks a question, the system performs a similarity search in MongoDB to find the most relevant context (text or visuals).
4. Generation: The configured Groq model generates an answer from the retrieved context, and Python appends the retrieved filename/page citations.

⚙️ Setup Instructions

1. Clone the repo:

   git clone https://github.com/amkr6207/Multimodal-RAG-Chatbot.git
   cd Multimodal-RAG-Chatbot

2. Create and Activate Environment:

   conda create -n chatbot-env python=3.12 -y
   conda activate chatbot-env

3. Install Dependencies:

   pip install -r requirements.txt

4. Configure Environment Variables: Create a .env file based on .env.example and add your:
   • GROQ_API_KEY
   • GROQ_VISION_MODEL (defaults to qwen/qwen3.6-27b)
   • GROQ_GENERATION_MODEL (defaults to qwen/qwen3.6-27b)
   • MONGODB_ATLAS_CLUSTER_URI
5. Run the App:

   streamlit run app.py

Vision Ingestion Behavior

Image captioning uses GROQ_VISION_MODEL. If one or more images cannot be captioned, the application stores the available text and successful image captions, then reports a partial-ingestion warning instead of silently claiming complete multimodal success.

Groq-hosted model availability can change. Check the Groq vision documentation before changing the configured model.

Tests

The automated tests use mocks and do not call Groq, MongoDB, or Hugging Face services.

pip install -r requirements-dev.txt
pytest

At startup, the application validates that GROQ_API_KEY and MONGODB_ATLAS_CLUSTER_URI are present and that the MongoDB URI has a valid scheme. This validation does not make external network requests.

Current Limitations

• Image understanding is caption-based; the original images are not passed to the final answer model.
• Citations identify retrieved chunks; they are not claim-level attribution from the generation model.
• Provider availability, rate limits, and network failures can cause partial image ingestion.
• The project does not yet include a RAG quality or latency benchmark.