Limits and generated-only evaluation¶
Use these bounds and evidence limits to decide whether assisted suggestions fit your document-review workflow.
Admission bounds¶
| Item | Limit |
|---|---|
| Complete envelope | 256 KiB |
| Records | 1–32 |
| Text per record | 16,000 Unicode code points |
| Complete encoder input | 256 tokens including query prefix and special tokens |
The service rejects overflow atomically and never silently truncates. The token limit is not a promise to cover a complete document of any length. Whole-document coverage is unknown. Provide already extracted text from one supported document; no PDF, image, audio or URL fetching is provided.
Seven supported types¶
Purchase order, delivery note, bank statement, utility bill, insurance policy, invoice and other business correspondence. Unknown, combined or insufficient text can still receive two plausible-looking suggestions. The tool does not detect those conditions or verify authenticity. Do not use rank as confidence or automate payment, credit, insurance or routing decisions from it.
What the evaluation measures¶
All 8,331 project examples are fictional/model-authored, including dependent variants. The original head fits 3,924 TRAIN rows through 459 family means. Exposed development results influenced selection. A separate frozen generated final retained 324 rows, including 72 out-of-scope controls; those controls still receive suggestions and are not counted as detected.
On the 252 supported final examples, the provisional author's label appears among the two suggestions on 84/84 English, 84/84 Spanish and 84/84 Portuguese examples. These are generated-corpus counts, not measured customer accuracy.
Label uncertainty remains material. With Gemma reviewer labels on that same fixed cohort, top-two coverage is English 81/84, Spanish 76/84 and Portuguese 71/84. Qwen's label is covered on 84/84 in each language. All author/reviewer disagreements remain without human adjudication; no final rows were relabeled or excluded. The model card linked from the examples repository retains the full evidence and limits.
There are zero native documents and zero independent human acceptance labels. Real customer-document accuracy is unmeasured. Test your own authorized representative documents and preserve manual choice/rejection. Original full-scope mixed/unknown recall objectives are not complete under this assisted release.
Measured runtime scope¶
Private us-east-1 real-time and certified package batch checks reproduced all 324 exposed reference outputs without differences. That is serving portability, not a new quality evaluation.
For 252 supported generated texts of 68–125 complete tokens, sequential client latency on one ml.c5.xlarge was p50 69 ms and p95 123 ms, excluding warmups and without SDK retries. It does not measure maximum-length inputs, concurrency, sustained load, customer workloads or an SLA. The example notebook uses us-east-1; Marketplace regional availability and AWS capacity remain separate from this measured scope.
See usage and support before integration.
Last updated November 16, 2026 · Suggest an improvement