Multimodal RAG Project

This repository contains a full-stack Multimodal Retrieval-Augmented Generation (RAG) backend powered by FastAPI, PostgreSQL/pgvector, and Google Gemini integration.

Features

• Document Ingestion: Processes PDF files extracting text (and theoretically images) chunks.
• Vector Search: Converts text into embeddings (via gemini-embedding-001) and performs fast similarity searches using pgvector.
• Chat API: Uses context mapped from vector searches to answer user queries with gemini-2.5-flash.
• Frontend UI: Interactive Chat and Document uploading natively in index.html.

Working

Multimodel.RAG.mp4

Prerequisites

• Python 3.9+
• Docker and Docker Compose (to run the PostgreSQL + pgvector instance)
• A .env file containing your valid Google Gemini API Key:

  GEMINI_API_KEY=your_api_key_here

Getting Started

1. Install Python Dependencies: Make sure you use a virtual environment (venv).

   pip install -r requirements.txt

2. Start the Database: Use the included docker configuration to effortlessly boot a pgvector compatible PostgreSQL wrapper.

   docker-compose up -d

3. Start the FastAPI Backend: Run the backend locally using uvicorn.

   uvicorn main:app --reload

4. Access the Frontend: Open index.html in your web browser, or serve it securely via your preferred web server tool (e.g. VS Code's Live Server extension).

Recommended Folder Structure

When the repository grows, consider separating the structure logically to keep it organized:

• /src - Core backend logic like main.py, ingest.py, etc.
• /frontend - Contains UI files such as index.html and static assets.
• /tests - Scratchpad validation scripts such as test_dim.py.

Note: User-supplied pdfs and vector-mapped images are excluded by default via .gitignore.