TypeChat

What it does

• Schema engineering replaces prompt engineering: declare your intents as types, and TypeChat builds the prompt, validates the model reply against the schema, and repairs non-conforming output through further model interaction
• A final step summarizes the validated instance — without using an LLM — and confirms it aligns with user intent
• Schemas compose: extend a discriminated union to add intents, or use a meta-schema to route user input to sub-schemas
• TypeScript/JavaScript is the first-class path; Python and C#/.NET exist as separate implementations (TypeChat.net)

When to use it

Natural-language interfaces where model output must land in typed objects: sentiment or intent classification, a shopping-cart API driven by chat, or any app where you would otherwise hand-roll JSON validation and repair loops.

Getting started

npm install typechat

Python and C#/.NET are separate codebases: see python/README.md and microsoft/TypeChat.net for working from source. Example projects live in typescript/examples; documentation at microsoft.github.io/TypeChat/.

Things to know

The README's PyPI and NuGet install lines are commented out: Python and C#/.NET are documented as work-from-source paths, not first-class package installs. TypeChat validates structure, not truth — schema-conformant output can still be factually wrong, so keep domain checks downstream.