Cookbook¶

These examples combine Kedi features into reviewable programs. Read the focused language pages first when you need a complete rule rather than a guided scenario.

Choose an Example¶

Example	Main concepts	Requires model calls
Structured Extraction	custom types, typed captures, native returns	yes
Tools and Approvals	Kedi and Python tools, risk, argument editing	yes
Agent Delegation	profiles, structured children, background lifecycle	yes
Evaluation and Optimization	datasets, metrics, `> optimize:`, GEPA	yes
Modules and Packaging	exports, selective imports, `package.kedi`	no
Complete Program	a compact end-to-end application	yes

Follow a Workflow¶

These tutorials keep their full listings beside the authoritative feature documentation. They are part of the cookbook, not additional APIs to learn.

Task	Start here	What to inspect
Transform nested collections	Loops and Map	collected native values, nested scope, conditional filtering
Route a generated draft using a scored judgment	Jev Workflow	generative model versus evaluator, threshold, both routing outcomes
Delegate, review, and write a report	Reviewed Evidence	typed child envelope, actual tool effects, edited approval, hooks
Compose child work dynamically	Dynamic variant	declared children, returned payload, separate approved writer
Find evidence in a large tool result	Artifact Retrieval	opt-in query, bounded reads, what remains outside model context
Join several large results	Artifact Reduction	executable reduction, bounded output, retained artifact handles
Embed Kedi in a Python application	Python Embedding	native result, runtime lifetime, cleanup
Validate before optimizing	Validation Workflow	exact outputs, failing cases, row errors versus suite errors
Develop across files	Local Package	explicit exports, isolated installation, receipt
Explore incrementally	Notebook	manual cell order, retained state, rerunning a cell

For deterministic and model-judged branch conditions, use Control Flow. A claim-driven branch is a model boundary; a Python predicate is not. Do not call a model for a condition your program can compute exactly.

Runnable Conventions¶

Unless a page shows a directory tree, save its complete listing as program.kedi and run:

kedi program.kedi

The model-backed listings in this section use openai:gpt-5.6-luna to make backend selection explicit. Replace it with a model configured for your environment. Provider credentials are read by the selected adapter; never place API keys in a .kedi source file.

Syntax Used in Examples¶

<value> substitutes an existing value into prompt or return text.
<python_expression> evaluates Python and renders its result as text.
[field: Type] in a >> block asks the model for a typed output.
[name: Type] = expression performs deterministic variable initialization.
= `python_expression` returns the native Python value.
[text] << prompt captures the raw model text; bare << is not an operator.

These distinctions matter. Use a typed output when a model must infer a value, Python when the answer is deterministic, and a native Python return when the caller should receive an object rather than its string representation.

Verification¶

Documentation CI parser-checks every kedi fence. Package manifests are parsed as package.kedi; a fence with a different required filename can declare file=..., and an intentionally non-parseable fragment must declare no-parse. This proves syntax and AST construction only. Imports, compilation, provider schemas, credentials, and adapter capabilities still require their documented file context and runtime tests against the production backend.

The offline documentation suite also executes the cookbook's complete programs: local package installation, real tool calls behind a local Pydantic FunctionModel, typed child delegation, and deterministic eval fixtures. The fixture supplies model responses; the runtime, tools, validators, and file effects are real. This proves those execution contracts, not provider accuracy, latency, or optimizer gains. The linked Jev tutorial checks routing with a test adapter, not live Jev scoring; artifact examples are model-assisted recipes, not measured retrieval-quality results.

In This Section¶

Pages