FAQ

Quick, honest answers to the questions people ask when evaluating KAOS. For the bigger picture, see what is KAOS? and how it fits together.

Do I need an API key?

Not for the golden path. Every tutorial and most how-tos run offline — deterministic packages run as themselves, and LLM examples use a built-in fake model (the FunctionClient seam). You only need a provider key for the optional live examples, which are clearly badged. Set KAOS_LEARN_LIVE=1 plus your key to switch an example to a real model.

Which packages do I actually need?

As few as one layer. The packages are deliberately small and composable — adopt kaos-pdf just to parse PDFs, or kaos-citations just to parse citations, without pulling in agents or LLMs. Dependencies point downward through the layers, never up.

Is KAOS production-ready?

It’s early — the packages are at v0 (0.1.x; a few are alpha), and APIs are still stabilizing. The substrate (documents, citations, NLP, retrieval, tabular) is solid and deterministic; the agent and LLM-programming layers are moving faster. Pin versions, read the changes & deprecations page, and expect some churn. This site only documents what’s real and tested today.

How is this different from LangChain / LlamaIndex?

Different bets, not a drop-in replacement. KAOS emphasizes:

• Typed programs over string-prompting — signatures and programs give validated, typed results, and an optimizer tunes them against a metric.
• Grounding and refusal as first-class outcomes — answers carry verifiable citations and refuse rather than hallucinate. This matters most for legal/financial work.
• A deterministic, offline substrate — BM25, citations, NLP, graph, and ML run without a model or a network, so they’re fast, cheap, and testable.
• MCP-native and offline-testable — every tool is serveable over MCP, and the whole stack runs deterministically in CI.

Can AI agents use KAOS directly?

Yes — that’s a design goal. Every package’s tools are MCP-native and can be served to a client like Claude Code or Codex. These docs are also published as /llms.txt and /llms-full.txt for one-shot ingestion.

Why “legal & financial”?

Because that’s where provenance, citations, and refusal are non-negotiable — a wrong, confidently-stated answer gets relied upon. The substrate is general-purpose; the design priorities are tuned for high-stakes document work.

What are the requirements?

Python 3.13+ and uv — that’s the whole setup for running examples. uv reads each example’s PEP 723 header and runs it in a cached, isolated environment.

Is everything on PyPI?

Most kaos-* packages are published to PyPI and install normally. A few app-stack pieces may need installing from source until everything lands — the capstone and the relevant examples note where this applies.

What’s the license?

Apache 2.0 — open source, published under github.com/273v by 273 Ventures.