> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cortex-labs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Develop and evaluate an enrichment

> Test a prompt and output schema on your own representative inputs before publishing a reusable version.

This exercise develops one small extraction. Choose a question and permitted dataset fragments that you can assess yourself.

You need Vantage, Forge authoring, sample access and an eligible model. Define success before testing.

<Steps>
  <Step title="Select representative cases">
    Choose a few clear positive, negative and ambiguous examples from your own material. Include a difficult or incomplete input.

    Record expected fields and the conditions for missing or uncertain values. Set quality, input and usage limits. An existing enrichment is optional.
  </Step>

  <Step title="Draft the definition">
    Open **Forge → Enrichments** and create the definition. Select supported fragment kinds and a permitted model.

    Write a focused extraction prompt and output schema. Use reusable schema blocks where appropriate. Require evidence locators for claims that need source support.

    For assistance, follow [Develop and test an enrichment with Kato](/kato/develop-enrichment).
  </Step>

  <Step title="Test in the playground">
    Select the authorised samples and limits. Run the candidate and inspect every output, schema failure and unsuccessful case.

    Keep the tested candidate and input versions identifiable. Playground output stays separate from current dataset evidence.
  </Step>

  <Step title="Assess and revise">
    Compare the outputs with your recorded criteria. Check omissions, unsupported fields, uncertainty and resource use.

    If an incumbent exists, test it on the same cases. Agreement with another model does not establish correctness.

    Revise the prompt or schema and repeat. Preserve failed cases in the evaluation record.
  </Step>

  <Step title="Publish the selected version">
    Record the evaluation decision and intended visibility. Publish the chosen immutable version.

    Deployment-wide publication requires evaluation evidence meeting the agreed criteria. Publishing does not expose private samples or activate datasets.
  </Step>
</Steps>

## Check the outcome

You should have an identifiable published definition version and evaluation evidence for your cases. Record remaining weaknesses before wider use.

Apply the version through [Apply an enrichment to a dataset](/vantage/bring-in-data/apply-enrichment). Binding, automatic processing and manual runs are separate decisions.

For deterministic definitions, evaluate the complete declared input set and reproducibility. Binding a deterministic bundle starts production work immediately, so use the intended destination.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.