Choose a model for JSON extraction

Define a passing answer

Require exact fields and types. Check current facts against historical distractors. Missing information must remain null, not a guessed value. Test extra fields and Markdown fences explicitly.

{
  "name": "林桐",
  "age": 28,
  "city": "杭州",
  "email": null
}

Compare the same workload

Use identical prompts, endpoint settings and output limits. Keep failed attempts in the denominator. Record the rate of fully correct records, first-content latency, completion time and billed tokens.

A faster wrong answer may cost more after retries or manual repair. Compare expected cost per accepted record only after measuring both failure rate and real billing usage.

Try a reproducible case

从以下资料提取联系信息。只输出一个 JSON 对象，恰好包含 name、age、city、email 四个字段；age 为数字，未提供的 email 为 null。不要 Markdown 或解释。 资料：林桐，28 岁，目前居住在杭州。未提供电子邮件。历史备注提到曾在北京出差，不是现居城市。

These synthetic contacts test exact fields and missing values. They do not establish performance on noisy PDFs, handwriting, multilingual records or your production distribution.