RAG Backend with Gemini Embedding + ChromaDB

This project is a lightweight Retrieval-Augmented Generation (RAG) backend using Google's Gemini embedding model and ChromaDB for vector search. It allows querying over static PDF documents stored in a local knowledge base (context_docs/) and occassionally uses Google Search Tool to amplify answers as well.

🗂️ Project Structure

.
├── main.py                # FastAPI app entry point
├── .env                   # Contains Gemini API key
├── requirements.txt       # Dependencies
├── context_docs/          # Static PDFs used as the knowledge base
│   └── tdr2024_en.pdf
├── chroma_db/             # Auto-created ChromaDB persistence
├── utils/
│   ├── loader.py          # PDF text extractor
│   ├── splitter.py        # Text chunk splitter
│   ├── prompter.py        # Generate prompt for the RAG AI Model
│   ├── embedder.py        # Gemini embedder logic
│   ├── vector_store.py    # ChromaDB logic for vector search
│   └── populate_db.py     # One-time DB population script

⚙️ Setup Instructions

1. Install dependencies

pip install -r requirements.txt

2. Set your Gemini API key

Create a .env file in the root:

GOOGLE_API_KEY=your_google_api_key_here

3. Add documents to context

Place your .pdf files in the context_docs/ folder.

4. Populate the Chroma vector store

python utils/populate_db.py

5. Run the API server

uvicorn main:app --reload

📫 API Usage

Once running, test the /query endpoint using Postman or cURL:

• POST /query
• Body (JSON):

{
  "query": "What are the major trade trends in 2024?"
}

• Response:

{
  "query": "...",               # This is the user query from API call
  "response": "...",            # This is the response after prompting the Gemini Model
  "grounding_metadata": "...",  # This is the grounding metadata if Google Search Tool is used (null if Google Search is not used)
  "context_used": "..."         # This is the documents used from vector db
}

✅ Notes

• You can update or add documents by dropping new PDFs in context_docs/ and re-running populate_db.py for the database name of trade_exports_db.
• Documents are indexed once with metadata for traceability.
• Sample docs for trade and export context has been added and indexed to the chroma db
• There are safeguards in-place to prevent the AI from answering any harmful queries. You can test this feature by asking: "Recommend me how to destroy my business competition in international trade and exports if I live in Indonesia"