article --summary
The summary
The article argues that giving agents live web search solves only half the problem. Once fresh pages arrive, the agent still has to decide which web results and which internal documents to rely on, and that decision belongs in the retrieval layer.
It describes how web results and internal company data differ in freshness, quality, structure, authority, and access control. Naively appending one result list to another tends to leave conflicts unresolved, let stale content win on similarity, apply permissions too late, and inflate cost and latency across many retrieval calls.
The author then covers design considerations: combining keyword, vector, and structured-filter retrieval in one query, and ranking on explicit freshness and authority signals using multiphase ranking. Examples from Clarm, Perplexity, and Metal AI illustrate these points.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.