SCHEMA = {

"申请人姓名": {"type": "str", "required": True},

"被申请人公司": {"type": "str", "required": True},

"入职日期": {"type": "date", "required": True},

"离职日期": {"type": "date", "required": False},

"欠薪金额": {"type": "money", "required": True},

"欠薪区间": {"type": "date_range", "required": True},

"事实与理由": {"type": "text", "required": True},

}

1. 模型只做抽取,输出 JSON,不做任何"补全想象"

extracted = llm_gateway.chat(

model="light", # 简单字段用轻量档

prompt=EXTRACT_PROMPT.render(schema=SCHEMA, user_input=desc),

response_format="json",

)

陈述类字段单独用强模型生成

extracted["事实与理由"] = llm_gateway.chat(

model="strong",

prompt=NARRATIVE_PROMPT.render(facts=extracted),

)

2. 确定性校验:规则代码兜底,模型不碰

missing = validate(extracted, SCHEMA) # 必填缺失、日期逻辑、金额格式

if missing:

return followup_questions(missing) # 缺什么问什么,不猜

3. 代码填充模板,原文条款一个字符不动

return render_template("劳动仲裁申请书.docx", extracted)

After launching the rebuilt version, mid-process drop-off dropped significantly. The most direct reason: users no longer experienced the "surprise of wrongly filled fields"—instead they got "follow-up questions about what's missing." The experience went from unpredictable to expected.

## Final Thoughts

The biggest lesson this project taught me: in AI applications, the model's boundary should be drawn smaller than you think. Let the model only do understanding and narrative generation, leave validation and assembly to code, and hand multi-model management over to a gateway—only then can a small team truly move fast and sleep well.

If you're also building a similar multi-model AI application and don't want to reinvent the wheel on key management, model routing, and failure fallback, try a unified AI API gateway: https://api.thistoken.ai/register

---

Tired of juggling provider integrations? Register at https://api.thistoken.ai/register and call every model through one base_url.