web search

In depth

The mechanism

The mechanism is that of any tool call. The model formulates a query, the harness runs it against a search engine, retrieves extracts or whole pages, and places them in the context. The answer is then produced from this enriched context. Two steps therefore determine the quality, and neither is the model’s doing alone: the formulation of the query, which decides what will be found, and the selection of what is retained, which decides what it will read.

Depending on a public index

This dependence on a public index has consequences that are underestimated. The quality of the answer inherits that of the web on the subject: on a niche question, the first pages can be promotional or automatically generated content. Sources contradict each other, and the model has no solid criterion for deciding: it will tend to follow the most confident wording. The best use therefore consists in requiring citations, and in treating the answer as a review of sources rather than as a verdict.

With RAG, not against it

Web search and RAG combine more often than they oppose each other. A serious system first queries the internal corpus, which is authoritative on procedures and in-house data, and opens the web only for what is external and recent: a market rate, a regulatory text, a news item. Telling them apart explicitly in the answer is good practice, because a reader does not treat a sentence from the internal manual in the same way as a sentence from some website.