An agentic approach to document search. Ask questions about your files and let the agent figure out what to read.

1. You ask a question
2. The agent lists files, reads relevant ones, searches semantically
3. It loops until it has enough context
4. It synthesizes an answer

git clone https://github.com/thirtyninetythree/soma-agent
cd soma-agent
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

Copy .env.example to .env and add your API key:

cp .env.example .env

python main.py --directory /path/to/your/docs

Options:

• --provider - claude (default), gemini, openai
• --directory - folder to index (default: current folder)

The agent has three tools:

• list_files - list directory contents
• read_file - read file contents
• semantic_search - search indexed documents by meaning

.txt, .md, .py, .js, .ts, .json, .csv, .html, .css, .yml, .yaml

On startup, the agent:

1. Walks the directory for supported files
2. Chunks each file (500 chars, 100 overlap)
3. Embeds chunks using all-MiniLM-L6-v2 (~90MB model)
4. Saves vectors to .rag_index/

Changed files are re-indexed automatically (hash comparison).

MIT