Skip to content
Learn KAOS

Compatibility & caveats

Python & platforms

Section titled “Python & platforms”
  • Python 3.13+ is required by every kaos-* package.
  • The Rust-backed packages (kaos-nlp-core, kaos-graph, kaos-ml-core, kaos-nlp-transformers) ship prebuilt abi3 wheels — no Rust toolchain needed to install or use them.
  • All packages are Apache-2.0 licensed.

Versions

Section titled “Versions”

Live versions are on the package reference (PyPI badges). For the authoritative version of any package:

Terminal window
uv run --with kaos-core python -c "import kaos_core; print(kaos_core.__version__)"

The packages are pre-1.0 and move quickly; pin versions in production (uv add 'kaos-core>=0.1,<0.2').

LLM provider caveats

Section titled “LLM provider caveats”

These matter when you run the live path (the offline FunctionClient path is unaffected):

  • Anthropic Haiku is the documented default for examples on this site — it’s cheap and has reliable cost accounting.
  • OpenAI reasoning models are not supported for the findings pattern — they’re incompatible with the temperature=0 the pattern needs, and some report cost as $0 (see cost as a contract on the honest cost gaps).
  • Set provider keys via ANTHROPIC_API_KEY / OPENAI_API_KEY / GOOGLE_API_KEY, or the KAOS_LLM_* equivalents (which take precedence). Keys are held in SecretStr and redacted from logs and output.

On-demand downloads & network

Section titled “On-demand downloads & network”
  • kaos-nlp-transformers downloads ONNX models on first use; pre-warm with kaos-nlp-transformers prefetch. The vendored potion-base-8M (via the [model2vec] extra) loads with no download.
  • kaos-web browser automation needs uv run playwright install chromium; HTTP extraction works without it.
  • kaos-source REST connectors hit live government APIs (EDGAR requires a User-Agent containing an email; GovInfo needs an API key; Federal Register / eCFR need neither).

Maturity note

Section titled “Maturity note”

kaos-ml-core is at an early (v0) stage: its CLI and serve entry points are stubs and the Rust crate exposes only a version probe. Use its Python API and MCP tools as the real surface; treat the CLI as not-yet-implemented.