LLM Extract API
Map text into supported fields while treating source instructions as untrusted content.
A focused use case
Classify permitted support messages into a controlled issue vocabulary and retain the passages supporting each requested field.
A practical workflow
Write a narrow task, separate source content from instructions, and require evidence for consequential values. Validate the schema and meaning outside the model; route absent or conflicting information for review.
An example record contract
These fields illustrate the decisions to make before implementation. Adapt the names and required values to the receiving system; the table is not a promise of a live endpoint or a provider-specific response.
| Example field | Meaning in this contract |
|---|---|
| source_id | Text identity |
| field_value | Candidate extracted value |
| evidence_span | Supporting passage |
| field_status | Supported, missing, or review |
Limits worth keeping visible
Prompt wording is not a complete security boundary. Keep secrets out of context and do not give an extraction-only worker permission to perform unrelated actions.
Before increasing the workload
Run a small, authorized test set through the whole path, including the receiving application. Keep accepted, rejected, and incomplete outcomes distinguishable. A result should retain the context needed to explain its meaning after it leaves the extractor.
Continue with LLM Extract API: Evidence Before Fluency for the full implementation discussion, and use the sample contract documentation to review the shape of a handoff.
Consult the source documentation for implementation details and current access conditions. The workflow above is an editorial planning pattern.