Prove · Structured output
Structured output
Define a JSON schema, run a prompt against it, and see either the validated output or the exact field that failed. Getting a model to return usable JSON is a skill you otherwise rebuild in every project. This is the same schema validation RepoOps runs on itself, turned into a surface you can use.
Where to find it
- Localhost:
/structured-output.html - API:
POST /api/structured-output/run,POST /api/structured-output/repair,GETandPOST /api/structured-output/schemas,POST /api/structured-output/schemas/delete,POST /api/structured-output/to-eval-bar - Sidebar: Prove, then Loop Engineering, then Structured output
What it does for you
/items/0/name points at the offending field rather than leaving you to diff two blobs. A model that returns malformed or non-conforming JSON reads invalid with those errors attached; it never reads as a quiet pass.Built vs. planned
All five routes ship today, backed by lib/routes/structured-output.mjs. The validator is dependency-free by decision: a small JSON-schema subset covering type, required, properties, items, enum, minimum and maximum, minLength and maxLength, and additionalProperties: false. That subset covers the structured-output use and keeps the repo's five-dependency budget intact, so a schema using a keyword outside it is not enforced. The pure pieces (validateAgainstSchema, evaluateStructuredOutput) are exported and tested without the network.
Last updated