ChatGPT Search Grounding and Reranking
The pipeline that runs when a question triggers retrieval: rewrite the query into search terms, query the underlying index, fetch page fragments, rerank them for relevance, and place the survivors in the context window to compose the answer.
Analogy Mental model
A secretary who knows the subject but does not trust her memory of this year's tax rules, so she calls the archive, has ten pages faxed over, picks the three most relevant, and writes the reply while reading them.
In practice What you actually do
Not every question triggers retrieval. General knowledge, code, and creative requests are answered from parameters with no sources attached. Timeliness, commercial comparison, local service, and fact-checking intent open the search path. Watching the outgoing query field in the browser's network panel shows exactly how your question was rewritten.
Where people get burned Pitfalls
That observation is one of the most useful techniques available: the terms the system chose to search with are the terms your page should be written around, and almost nobody looks at them.
This entry was written in English from the Chinese original, not machine-translated. Read the Chinese version →