Usage¶
Supply explicit source segments and inspect every generated draft against the quoted source. Keep segment IDs stable within the submitted pool.
Request¶
{"language":"en","segments":[{"id":"s1","text":"Maya Chen will send the signed plan by 2027-06-18."}]}
Set language to en, es or pt and provide text in that language. Every segment needs a unique nonempty id and nonempty text. Use application/json, or application/jsonlines with one complete structured record per invocation. Batch input must split on lines.
| Bound | Limit |
|---|---|
| Segments | 1–64 |
| Text per segment | 2,000 Unicode code points |
| Combined segment IDs and text | 16,000 code points |
| HTTP body | 96 KiB |
| Actual model input | 3,072 tokens |
| Generated output | At most 1,024 tokens |
Oversized inputs reject without silent truncation. This interface accepts text, not PDF, images or audio; prepare extracted text separately.
Response and evidence¶
The response contains model identity, language, summary, actions, evidence_check and limitations. Each draft item includes an evidence object with segment_id, quote, start and end. The quote is the full selected source segment, copied by the server. Offsets are Unicode code points: start=0 and end equals that segment's code-point length.
The evidence strategy locates a literal source segment; it does not prove that the draft follows from it. review_required is always true. Inspect the original text, rather than treating a copied quote as a factuality guarantee.
Action owner needs an explicitly named individual from the source. deadline needs an explicit valid YYYY-MM-DD calendar date. Unknown values remain null; relative phrases such as “next week” are not converted to dates. Review all candidates before creating tasks in another system.
Last updated November 15, 2026 · Suggest an improvement