Installation

ragX needs Python 3.10 or newer.

pip install pyragx

The distribution is named pyragx on PyPI; the package you import and the command you run are both ragx.

The base install runs the default variant, ragx-lexical, which ranks with BM25 and needs no embedding model, no vector database and no GPU.

Optional extras

Install extras for the parts you use:

Extra	Install	What it adds
dense	pip install "pyragx[dense]"	sentence-transformers and NumPy, for the dense, fused and rerank rankers (ragx, ragx-core, ragx-fused, ragx-rerank, ragx-fast, ragx-think)
anthropic	pip install "pyragx[anthropic]"	the Anthropic SDK, for anthropic: models
google	pip install "pyragx[google]"	the Google Generative AI SDK, for google: models
local	pip install "pyragx[local]"	transformers, PyTorch and Accelerate, to run hf: models in-process
server	pip install "pyragx[server]"	FastAPI and Uvicorn, for ragx serve
ui	pip install "pyragx[ui]"	Textual, for the full-screen terminal view
all	pip install "pyragx[all]"	everything above

OpenAI-compatible providers (OpenAI, Groq, Together, Cerebras, Fireworks and any vLLM server) work with the base install.

Set a model

ragX reads provider keys from the environment, or from a .env file in the working directory or any parent. It never logs them.

export OPENAI_API_KEY=...        # or ANTHROPIC_API_KEY, GOOGLE_API_KEY, GROQ_API_KEY, ...

Without model=, ragX picks a default from the first key it finds (see Models and providers). With no key at all it falls back to a small local model, which needs the local extra.

Optional tools

The classic agent uses ripgrep for exact search when it is on the PATH (and falls back to a pure-Python search when it is not), and ugrep for fuzzy search. The ragX variants need neither.

Verify

ragx --version
ragx variants