Usage¶
Send ticket text to the endpoint or to a batch transform job. Keep a copy of the ticket id in the JSON when you need it back. JSON and JSON Lines preserve string ids. Plain text is one ticket and has no id.
Requests¶
Content-Type: application/json
{"id": "T-1", "subject": "Cobro duplicado", "body": "Me cobraron dos veces. Revisen la factura hoy."}
{"text": "..."} classifies that string. When both are present, text
wins over subject and body.
| Input | What it means |
|---|---|
| JSON object | One ticket, as above |
JSON array, or {"instances": [...]} |
Several tickets. The response is {"predictions": [...]} |
| JSON Lines | One ticket per line, and one prediction per output line |
| Plain text | One ticket |
| CSV | The text column, or the first column when there is no header |
Requests are UTF-8, at most 6 MiB, and at most 32 records. Each ticket is
at most 100,000 characters. Blank text is rejected. The encoder reads the
first 128 tokenizer tokens, including its prefix and special tokens. Text
after that window can be dropped, and truncated is then true.
Batch configuration on the listing is SingleRecord, SplitType Line,
up to 6 MiB. The server admits two inference calls at the same time.
Result fields¶
| Field | Meaning |
|---|---|
category, intent |
The same predicted category, with a score, a mode, and a ranked distribution |
top_k |
The top three category candidates, or fewer for a restricted label set |
urgency |
low, medium, or high |
sentiment |
negative, neutral, or positive |
language |
es, pt, en, or unknown, plus a score. Language confidence is uncalibrated |
needs_review, review_reasons |
A review signal for uncertain language or category, short input, multiple requests, conflicting clauses, truncation, or custom labels |
truncated |
True when input past the 128-token window was discarded |
model, version, model_revision |
Model name, version, and a SHA-256 fingerprint of the runtime sources, metadata, and classifier heads |
A false review flag does not mean the prediction is correct. An empty
review_reasons array is not an accuracy guarantee. Route tickets you
would act on through a person, including the ones with needs_review
false.
The published real-time sample on the listing, for an English billing
sentence, returns billing at 0.5518, negative sentiment at 0.9251,
high urgency at 0.5473, and language en. Those scores are the published
example for that sentence. They are not calibrated probabilities, and a
different sentence gets a different result.
Errors¶
| HTTP status | What to do |
|---|---|
| 400 | Fix malformed input, blank text, labels, or the record limit |
| 408 | Retry a timed-out body upload |
| 413 | Reduce the payload below 6 MiB |
| 415 | Use a supported content type |
| 429 | The two inference slots are busy. Retry with backoff and jitter, and lower concurrency |
| 500 | Check readiness. Write to support if it persists |
The server does not log ticket bodies or query strings. You control AWS access logs, data capture, and what you store downstream.
Custom labels¶
This path is experimental. A request can pass
{"text": "...", "labels": {"intent": ["wifi outage", "picnic planning"]}}.
A known subset uses mode finetuned_subset. Novel labels use zero_shot.
The default taxonomy uses finetuned. Category overrides trigger review.
The published quality figures apply to the default labels only. Do not
compare a restricted-set score with a full-taxonomy score.
Last updated November 11, 2026 · Suggest an improvement