These are the workflows KAOS was built for — the same use cases the legacy Kelvin
documentation covered, rebuilt fresh on the new packages. Every one has a runnable,
tested example you can run in one command. They use small inline samples so they run
offline; the Real data note on each points at the production data source — including the
public kl3m datasets on Hugging Face.
Most reduce to three KAOS patterns: classify (typed Call → a label), extract
(typed Call → a structured record), and review/answer (an agent over a corpus). That’s
the point — a handful of composable primitives cover the whole domain.
flowchart LR
doc["📄 Document"] --> classify["<b>Classify</b><br/><small>typed Call → label</small>"]
doc --> extract["<b>Extract</b><br/><small>typed Call → record</small>"]
doc --> review["<b>Review / Answer</b><br/><small>agent over a corpus</small>"]
classify --> c["triage · UTBMS coding · matter tagging"]
extract --> e["complaints · covenants · deal terms · abstraction"]
review --> r["contract review · due diligence · research"]
classDef p fill:#eef2ff,stroke:#6366f1,color:#1e1b4b;
class classify,extract,review p;