The offline seam

Most documentation for LLM tooling can’t run its own examples: they need an API key, cost money, and give different output every time. So the examples rot, quietly, until a reader copies one that no longer works.

Learn KAOS takes a different stance: every example runs offline, for free, and is tested in CI. The mechanism is a single design choice in kaos-llm-client.

One interface, two implementations

kaos-llm-client exposes every provider — Anthropic, OpenAI, Google — through one interface (client.chat(...), structured output, streaming, cost accounting). It also ships a FunctionClient: a client that runs a Python callable instead of making an HTTP request, while satisfying that same interface.

flowchart LR
    code["Your code<br/><small>client.chat(...)</small>"] --> iface{{"kaos-llm-client<br/>interface"}}
    iface -->|real| prov["Provider client<br/><small>Anthropic · OpenAI · Google</small>"]
    iface -->|fake| fc["FunctionClient<br/><small>a Python callable</small>"]
    prov --> http["🌐 HTTP · key · $ · nondeterministic"]
    fc --> det["✅ offline · free · deterministic"]

    classDef seam fill:#eef2ff,stroke:#6366f1,color:#1e1b4b;
    classDef good fill:#f0fdf4,stroke:#22c55e,color:#14532d;
    class iface seam;
    class fc,det good;

So your code calls client.chat(...) identically whether the client is a real provider or a deterministic fake. Only the model changes:

client = FunctionClient(function=fake_model)   # offline, free, deterministic
# vs
client = create_client("anthropic:claude-haiku-4-5")   # live

This is the seam the FunctionClient tutorial teaches. Every LLM example uses a make_client() that returns the fake by default and the real one when KAOS_LEARN_LIVE=1 is set.

Why this matters for you

• The code you read is the code that’s tested. Pages import their example files byte-for-byte; CI runs those files on every change. A broken example fails the build.
• You can run everything with no account. Learn the whole spine before you ever pay for a token.
• Your own tests get faster and cheaper. The same FunctionClient makes your LLM code deterministic and offline in your test suite — see write a tested example.

Determinism beyond LLMs

Many KAOS packages are deterministic by nature — kaos-citations, kaos-graph, kaos-nlp-core, kaos-content, kaos-core. They need no seam at all: same input, same output, always. That’s why the gallery leads with them — the most convincing demo is one that runs identically on your machine and ours.