Custom Agent Framework for RAG

Table of Contents

• About
• Documentation
• Getting Started

🚀 About

This project is the base for a hand-made framework to build retrieval-augmented assistants.
It’s designed as a modular foundation: documents are ingested, searched, and reasoned over by agents that orchestrate specialized tools. The structure makes it easy to extend with new retrieval methods, models, or task-specific logic.

⚡ This is just the starting point: the codebase doubles as a lightweight, hand-crafted framework for building custom agentic workflows around documents and retrieval.

📚 Documentation

What It Does

1. Ingests documents

   • Scans a data/ folder for PDFs.
   • Extracts page-level text using PyMuPDF.
   • Creates a searchable collection with ChromaDB.

2. Retrieves relevant content

   • Queries the vector store with natural-language questions.
   • Surfaces the most relevant document passages, with metadata (source and page number).

3. Generates responses

   • Uses Google Gemini models to reason over retrieved passages.
   • Produces concise, context-grounded answers.

4. Coordinates reasoning with agents

   • An Orchestrator agent decides how to break down a query.
   • Specialized tools (retrieval, response generation, etc.) handle subtasks.
   • Agents maintain a short-term memory of queries and tool outputs, enabling iterative refinement.

Tech Stack

• Document parsing: PyMuPDF
• Vector DB: ChromaDB
• LLM API: Google Gemini
• Configuration: .env for API keys
• Framework: Agent + Tool abstraction using Pydantic models

📝 Getting Started

python main.py

Example Query:

Compare FDA and WHO approaches to risk management for medical devices

Example Flow:

1. Orchestrator agent receives the query.
2. Delegates retrieval to the RAG tool.
3. Retrieves relevant pages from FDA and WHO docs.
4. Passes results to the Response tool.
5. Outputs a concise, well-grounded comparison.

Back to top