Develop and test an enrichment with Kato

Use Kato to help develop an enrichment when you need structured information from your data. You supply the purpose and assessment criteria, inspect the outputs, and decide which version is suitable for reuse.

​

Define the task and sample

You need permission to author the definition, read the sample inputs, use the selected model and run an evaluation. Applying it later also needs the dataset, execution-principal and processing permissions. Choose a small representative sample of your own permitted inputs and the fields you need from them. Inspect each source and write the expected result before evaluation. Include a clear positive case, missing or ambiguous values, irrelevant content and conflicting wording. Use supported input kinds and retain the exact sampled versions; no sample dataset is supplied by the guide. Before running, set sample and usage limits. Define success: supported fields are accurate, missing values remain explicit, and the output invents no cause, actor or date. Retain each source version and your assessment criteria.

​

Ask for a draft

In Kato, send:

Help me draft an enrichment named [definition name]. It reads [my supported input kind] and returns [fields required for my task], with source support and uncertainty. Use null or an explicit unknown state for unsupported values. Do not invent missing facts. Propose the prompt and output schema, then leave the definition unpublished and unbound.

Replace the fields and supply your expected output shape before authoring. Select the permitted model and supported input kinds with Kato. Inspect the draft in Forge and its enrichment playground. Check that the prompt and schema agree on field names, types, missing values and how supporting text is retained. Change only the candidate you intend to test.

​

Evaluate the candidate

Identify the permitted sample scope and assessment criteria. Ask Kato to run an isolated evaluation of that candidate with the agreed limits, then open the evaluation record. Inspect actual outputs, schema errors, failed cases, quality judgements, resource usage and cost. Compare output with the source text and expected results. For a required value that is absent, check the explicit unknown result. For conflicting accounts, check that they remain distinguishable. Evaluation outputs stay separate from current dataset evidence and production processing. The sample and results keep their access rules and classification. Do not copy private example content into a definition intended for a wider audience.

​

Revise and repeat

Ask for a specific correction, for example:

Revise this candidate to address [the specific error I found]. Preserve unknown values when the source does not support them. Repeat the evaluation against the same input versions and criteria.

Keep the tested input and candidate versions with their results. When there is an incumbent, compare both on the same cases and criteria. Agreement with an earlier model does not establish correctness; assess each against the source and criteria. Failed examples remain part of the evaluation record.

​

Publish a selected version

Once the criteria are met, choose the exact candidate and intended visibility. Publish its immutable definition version explicitly, retaining the evaluation evidence and publication decision. Deployment-wide publication requires a recorded evaluation that meets the agreed criteria for that version. Kato can assist with the operation if you authorise it, but its recommendation does not decide evaluation success. Publication does not expose private evaluation records or update active dataset bindings.

​

Apply it deliberately

Choose the authorised dataset and the published definition version. Ask Kato to configure the dataset binding and show its execution principal and payer team. Confirm the scope: a dataset binding can be used by other workspaces that reach that dataset. The platform assigns the payer from the binding’s scope. Choose a bounded manual pass or explicitly enable automatic processing. Enabling automatic processing schedules eligible current and future inputs. A manual pass can run while automatic processing remains off; a paused binding must be resumed before it runs. Inspect the pending-item preview and supported limits before admission. Check the resulting work and a sample of dataset outputs against their input and definition versions. Repeating unchanged completed work reuses results; it need not make another model call. An explicit forced rerun records its purpose and incurs separate computation. Changed inputs or definitions require new processing while retained earlier evidence keeps its selected versions. Use Find resources and track work with Kato for progress and Recover an incomplete Kato task for failed or partial work. Key concepts distinguishes a definition, binding and result. For the direct authoring route, use Develop and evaluate an enrichment. For version binding, pass limits and deterministic processing, use Apply an enrichment to a dataset.