Home Original page

GitHub - leonardopra/python-agent-lab: RAG pipeline, multi-agent system and LLM tools built from scratch in Python

RAG pipeline, multi-agent system and LLM tools built from scratch in Python.

Overview

This repository documents a progressive build of AI agent systems using the Anthropic API and open-source tooling. Each module is self-contained and introduces one new concept on top of the previous.

Modules

File Description
hello_llm.py Basic API call to Claude — prompt in, response out
agent.py Stateful agent class with conversation history
agent_with_tools.py Tool-calling agent with calculator and note search
agent_with_memory.py Persistent memory via JSON file + structured logging
rag.py Minimal RAG pipeline with ChromaDB and sentence-transformers
rag_v2.py Improved RAG with better chunking and similarity filtering
multi_agent.py Multi-agent system with Planner routing to specialized workers

Architecture — multi_agent.py

User query
    │
    ▼
 Planner  ──── Claude API
    │
    ├──► RAGWorker   → ChromaDB → docs2/
    ├──► CalcWorker  → Python eval()
    └──► ChatWorker  → Claude API

The Planner reads the incoming query and routes it to the appropriate worker:

  • RAGWorker — retrieves relevant chunks from the vector store and generates a grounded answer
  • CalcWorker — evaluates mathematical expressions
  • ChatWorker — fallback for general conversation via LLM

Tech Stack

Component Technology
LLM Claude (claude-sonnet-4-6) via Anthropic API
Vector DB ChromaDB
Embedding model sentence-transformers
Framework Python 3.14 + LangChain
Environment python-dotenv

Setup

# Clone the repo
git clone https://github.com/YOUR_USERNAME/python-ai-agent-lab.git
cd python-ai-agent-lab

# Create and activate virtual environment
python -m venv venv
source venv/bin/activate  # macOS/Linux

# Install dependencies
pip install anthropic langchain langchain-anthropic chromadb sentence-transformers python-dotenv

# Add your API key
echo "ANTHROPIC_API_KEY=your_key_here" > .env

# Run any module
python multi_agent.py

Known Limitations

  • Planner routing depends on LLM text output — brittle for edge cases
  • ChromaDB returns n results regardless of relevance (no similarity threshold in v1)
  • No conversational memory across sessions in the multi-agent system
  • CalcWorker does not handle malformed expressions gracefully

Roadmap

  • Add similarity threshold filtering to RAGWorker
  • Persistent session memory across conversations
  • Streaming responses
  • REST API layer with FastAPI
  • Unit tests for each worker

License

MIT