The outcome

Label or score thousands of documents on many dimensions at a cost measured in dollars rather than hundreds, without a chain of dependent prompts.

Step by step

A workflow you can repeat.

  1. 01

    Write the full set of questions up front — the classification Choice, the relevance Nouls, the quality Scores — including speculative ones your pipeline only needs in some branches, because an extra question costs only its own tokens and almost no latency.

  2. 02

    Filter each record in code before it becomes state: trim to the fields the questions actually read, since accuracy drops as unrelated material grows and the limit is 64k tokens for state plus questions.

  3. 03

    Batch through the async SDK client with a concurrency cap under the published rate limits (1,200 requests per minute and 250,000 tokens per second at launch) and let the SDK's default retry policy handle 429 and 529 responses.

  4. 04

    Store the raw probabilities alongside each label so downstream analysis can re-threshold without re-running the corpus, and treat the score field as a position between your levels, not as a precise number.

  5. 05

    Sample the low-confidence tail for human review and use what you find to tighten the criteria wording, then pin the versioned model ID so a future jev-latest release cannot silently shift your labels.

Working standard

What good use looks like.

  • Filter before you send.
  • Pin the model version for reproducible labels.
  • Review the low-confidence tail, not a random sample.

Official references

Check the current product documentation.