Grounding answers with strict web-search gating

A search tool the model reaches for on every message is worse than no tool at all. It slows responses, burns budget, and — worst of all — invites the model to invent citations. The fix isn't a better search API. It's a better decision about whether to search in the first place.

Intent before retrieval

Before any tool runs, a classifier labels the request. Transform tasks — 'summarize this', 'convert to a document', 'rewrite this email' — never trigger a search. Real-time and factual-lookup queries — 'who won', 'latest', 'current price' — always do. Ambiguous cases fall back to context.

• Transform → never search; go straight to generation.
• Real-time / factual-lookup → always search.
• Informational → answer from the model unless a fact is missing.
• Ambiguous → search only when required entity knowledge is absent.

Query rewriting and ranking

When we do search, one query is rarely enough. We expand the request into several variants — entity plus location, quoted phrases, domain-specific forms — run them in parallel, dedupe URLs, and rank by exact match and source authority.

Confidence gating

Results get a confidence score. If multiple strong sources agree, we answer with citations. If the evidence is sparse or contradictory, we say so plainly instead of guessing. No fabricated links, no 'temporary' download URLs — if we can't verify it, we tell you.