Minimal RAG Agent

Made by Rodrigo de Sarasqueta

Ultra-clean RAG agent skeleton with multi-LLM support and dual memory.

Features

• Multi-LLM support (Ollama, OpenAI, DeepSeek)
• Dual memory system (buffer + summary)
• PDF/TXT document ingestion
• Semantic search with ChromaDB
• RESTful API endpoints
• Minimal CLI interface
• Docker deployment

Quick Start

# Install dependencies
pip install -r requirements.txt

# Add documents
python cli.py add-pdf document.pdf

# Start chat
python cli.py

# Or start API server
python api.py

# Evaluate performance (NEW!)
python evaluate_mock.py test_questions.json  # Basic test (5 questions)
python evaluate_mock.py test_questions_simple.json  # Simple test (8 questions)  
python evaluate_mock.py test_questions_comprehensive.json  # Comprehensive test (12 questions)
python evaluate_simple.py test_questions.json  # Simple evaluation with real LLM
python evaluate.py test_questions.json  # Full RAGAS evaluation

Configuration

Edit config.json:

{
  "llm": {
    "provider": "ollama",
    "ollama_url": "http://localhost:11434",
    "ollama_model": "llama3:latest"
  }
}

For API keys, use environment variables:

export OPENAI_API_KEY=your-key
export DEEPSEEK_API_KEY=your-key

Docker

docker-compose up

API Endpoints

• POST /chat - Chat with agent
• POST /upload - Upload document
• GET /documents - List documents
• DELETE /documents/{id} - Delete document
• GET / - Web interface

Architecture

Agent -> LLM Client (Ollama/OpenAI/DeepSeek)
      -> Vector Store (ChromaDB)
      -> Dual Memory (Buffer + Summary)
      -> Tools (RAG + HTTP)

Structure

├── core/
│   ├── agent.py         # Main agent
│   ├── llm_client.py    # Multi-LLM support
│   ├── vector_store.py  # ChromaDB RAG
│   └── tools/           # RAG and HTTP tools
├── api.py              # FastAPI server
├── cli.py              # Minimal CLI
├── config.py           # Configuration
└── config.json         # Settings

Clean foundation for building AI agents.