LLM Retrieval Augmented Generation (RAG) - Basic application with FastAPI, NextJS and MongoDB

A basic simulation of RAG system using MongoDB deployed by FastAPI and NextJS

Key concept:

Retrieval Augmented Generation (RAG)

There are 5 key components in a basic RAG:

1. Document Loader:
- PDF Loader
2. Document Splitter:
- Text Splitter
3. Vector Embeddings:
- Embedding model: HuggingFace Embedding Model - all-MiniLM-L6-v2-Q5_K_M.gguf
4. Vector Store
- MongoDB Atlas - Vector Stores
5. Retrieval and Generation
- Chat model: Hugging Face Chat Model - Phi-3-mini-4k-instruct-q4.gguf

Deployment

1. Backend: FastAPI
2. Frontend: NextJS

How to run?

!Note: The system need to be run with 2 components (backend and frontend) seperately:

• Run Backend in the first terminal window:
cd backend python main.py

!Important: Check installed libs in 'requirements.txt' and environment variables in '.env.example'

• Run Frontend in the second terminal window:
cd frontend npm run dev

Shortcommings

• The system is only allowed to upload only 1 document
• The system use gguf model file to correspond to CPU
• The system is seperated into 2 individual parts (backend and frontend) -> Docker can solve this!

What's next?

• Advanced RAG: RAG with external API