Retrieval

Retrieval starts with a user or system query and ends with a candidate set or context passages. The corpus may be a web index, a private database, or uploaded files. Direct observation requires traces, disclosed documentation, or a controlled system.

Section sources:[1] Google Search Central[2] OpenAI Developers

A crawler fetches pages, an index stores or organizes an accessible representation, ranking orders candidates, and retrieval selects them for a particular task. Citation and attribution belong to the visible answer. One event is not evidence of the others.

Section sources:[1] Google Search Central[2] OpenAI Developers

OAI-SearchBot may request a page, but a server log does not show whether it entered the candidate set for a later ChatGPT Search answer. Conversely, a link in an answer is an observable outcome but does not expose every document the system considered.

In an external audit, preserve identical queries, mode, language, region, answers, and final links; this measures outcomes compatible with retrieval, not the hidden stage itself. In your own system, also log the query, corpus version, candidate IDs, and scores.

A missing link does not prove missing retrieval because used context may be hidden. A present link does not reveal why it was selected. External studies should therefore report observable answers and links, not internal retrieval without trace access.