跳到主要内容Skip to main content
← 深水指南Guides

引用逐条核对Check citations one by one

提示词与工作流Prompts & workflows · 提示词案例Prompt case

输出断言-引用核对表;不支持或错位引用标红。

Output a claim-citation check table; unsupported or misplaced citations are flagged.

提示词效果未实测。Prompt effect has not been tested.

提示词是可复跑的模板,不是效果承诺。换模型、换版本、换输入都可能改变结果;用本页的检查点自己核对,并把官方出处当作机制依据而不是效果保证。A prompt is a re-runnable template, not a promise of results. Changing the model, version or input can change the outcome. Use the checkpoints on this page to verify it yourself, and treat the official sources as evidence of mechanism rather than a guarantee of effect.

任务与输入Task and inputs

核对每个断言是否有对应引用支持。 输入:带引用文本、引用来源片段。

Output a claim-citation check table; unsupported or misplaced citations are flagged. Inputs: Text with citations and source excerpts.

提示词本体The prompt

逐句核对文本。输出 checks[]:claim、citation、source_excerpt、verdict(supported/unsupported/misplaced)、reason。source_excerpt 必须来自给定来源片段;找不到对应片段则 verdict=unsupported。不得用常识补证。

占位符用 {字段} 标注,运行前替换为本次输入。

Check the text sentence by sentence. Output checks[] with claim, citation, source_excerpt, verdict (supported/unsupported/misplaced), and reason. source_excerpt must come from the supplied excerpts; if no matching excerpt exists, verdict=unsupported. Do not use common knowledge as evidence.

Replace {field} placeholders with the current inputs before running.

预期输出形态Expected output shape

checks[],每条含判定与理由。

checks[] with verdict and reason per claim.

检查点Checkpoints

  • 每个 verdict 有 source_excerpt 或明确缺失。
  • misplaced 指出正确来源应是什么。
  • 不引入外部常识。
  • Each verdict has a source_excerpt or explicit absence.
  • Misplaced citations identify the correct source.
  • No external common knowledge is introduced.

失败分支Failure branches

  • 来源片段不足:补片段再核。
  • 断言过宽:拆成可核对子句。
  • Insufficient excerpts: add excerpts and recheck.
  • Claim too broad: split into checkable subclaims.

出处与边界Sources and limits

提示词效果未实测。机制依据见出处;换模型、版本或输入会改变结果。

This site has not tested the effect of this prompt. See the sources for the mechanism; changing model, version or input can change the result.

出处Sources

来源Source链接URL
OpenAI Cookbook · Building resilient prompts using an evaluation flywheel(官方仓库 raw,2026-11-07 直抓 200)OpenAI Cookbook: Building resilient prompts using an evaluation flywheel (official repository raw, fetched 200 on 2026-11-07)https://raw.githubusercontent.com/openai/openai-cookbook/main/examples/evaluation/Building_resilient_prompts_using_an_evaluation_flywheel.mdhttps://raw.githubusercontent.com/openai/openai-cookbook/main/examples/evaluation/Building_resilient_prompts_using_an_evaluation_flywheel.md

资料快照日期:2026-11-07;单项复核范围与日期见对应正文或来源记录。产品功能与条款可能变化,使用前请核对原始来源。Source snapshot: 2026-11-07; item-specific rechecks and their scope are recorded in the text or source notes. Features and terms may change; check the original sources before use.