跳到主要内容Skip to main content

引用与可追溯Citations and Traceability

Citations and Traceability

  • 数据与检索Data & retrieval
  • 第三站 · 用深Station 3 · Going Deeper
  • 引用核验 2026-11-06refs checked 2026-11-06
解决什么问题What problem it solves

让回答里每句话能倒查回来源、版本和原文位置,而不只是看起来有出处。Make every sentence in an answer traceable to a source, a version and a location rather than merely looking sourced.

这是什么What it is

引用要看对应关系,不是看链接数量。ALCE 把带引用的生成拆成流畅度、正确性、引用质量三类自动指标,其中引用质量检查的是这句话能不能被它引的那份来源支持。

托管的接地会把结构直接给你。Google 搜索接地的官方文档描述:响应里带来源数组(含 uri 与 title),另有一个把回答文本片段对应到来源的结构,界面用它渲染内嵌的来源链接,并按规定展示提示。

媒体文件的出处走另一条路。C2PA 是一套开放标准,用凭证记录内容的来源与编辑过程,跟「回答引用网页」不是一回事:它管的是素材本身,不是文本里的引用。

版本与时间点也要留:模型版本、知识截止、检索到的文档版本、抓取日期。缺一样,倒查就只能停在「大概那个时候」。

Citations are about correspondence, not link count. ALCE defines automatic metrics across fluency, correctness and citation quality, where citation quality checks whether the statement holds up against the source it points to.

Hosted grounding hands you the structure directly. Google's documentation for search grounding describes a response containing source entries with a uri and a title, plus a mapping from spans of the answer text to those sources, which an interface uses to render inline citation links under the required display terms.

Media provenance runs on a different track. C2PA is an open standard for recording where a piece of content came from and how it was edited; that is about the asset itself, not about a text answer citing a page.

Keep versions and timestamps as well: model version, knowledge cutoff, the revision of the retrieved document, the date it was fetched. Miss one and the trail only goes back to sometime around then.

为什么需要知道Why it matters

对外发布的结论、涉及合规的判断、需要别人重做的实验,都得能倒查回去。能倒查的输出才算可复用,剩下的只是读起来顺。Published conclusions, compliance decisions and experiments someone else must repeat all need a way back to the source. Output you can trace is output you can reuse; the rest only reads smoothly.

出处Sources
  • Gao 等《Enabling Large Language Models to Generate Text with Citations》(ALCE,EMNLP 2023)ALCE: citation-quality metrics for generated text ↗
  • Grounding with Google Search:响应里的来源与文本片段对应关系Grounding with Google Search ↗
  • C2PA 官方:内容出处与编辑凭证的开放标准C2PA provenance standard ↗
  • OpenAI 网页搜索工具指南OpenAI Web Search tool guide ↗