Line-by-line Search · 文档级语义 grep

来源:docs.typesafe.ai/cookbooks/semantic_find 抓取时间:2026-09-20

模式

resp = client.system_one(
    state={"query": "What is the refund policy?", "lines": list_of_218_lines},
    questions={
        f"rel_{i}": Score(["irrelevant","tangential","relevant"],
                          instructions=f"How relevant is line {i}?")
        for i in range(218)
    } | {
        "contains_answer": Noul("Does this document contain the answer?")
    }
)

两道题共用一个 state:

• Score 题:对每行打分(精确但贵)
• Noul 题:对整篇判断是否有答案(便宜,先做 gate)

何时用

技术文档 / 法律合同 / 长篇 FAQ 的语义搜索。比 embedding 检索更精确,比 LLM 重排序便宜 100 倍。