| name | search-discipline |
| description | Use whenever deciding whether to search the WEB and how to construct the queries - answering factual questions, evaluating products/versions/APIs, encountering unfamiliar names or terms. Encodes the rate-of-change rule, the unrecognized-entity rule, effort scaling, query construction, and source-trust calibration. Do not use for pure reasoning tasks with no factual dependency, nor for lookups inside a local repo or docs tree - those are grep/read questions, not search questions. |
search-discipline: search on rate-of-change, not on confidence
The failure mode is never "searched unnecessarily" - a wasted search costs seconds. The failure mode is answering from stale training data with fluent confidence. Searching costs seconds; confabulating costs the user's trust. When in doubt, that asymmetry decides.
When to search
- Rate of change, not topic. Ask: how fast does the true answer change? Daily/monthly (prices, versions, releases, holders of positions, policies, availability) -> always search. Effectively never (math, established science, historical fact, language semantics) -> answer directly. In between -> answer from knowledge AND verify the volatile parts.
- The unrecognized-entity rule. An unfamiliar capitalized word in the request is almost certainly a name that postdates your training - not a common noun you can gloss over. The test: does answering require knowing what that thing is? If yes and you cannot place it: search. Knowing a franchise, vendor, or product line is NOT knowing its newest member.
- Specific product / model / version / recent technique named? Search before answering. Partial recognition from training does not mean current knowledge.
- Present-tense phrasings of "settled" questions ("does X exist", "is Y still the case", "who currently...") - the phrasing itself signals the answer may have moved. Search.
- Personal/company context ("our pipeline", "my repo", company-specific terms) -> look in the local/internal source first, the web second. Possessives are routing signals.
How to search
- Short queries win. 1-6 words. Start broad (1-2 words), then narrow with follow-ups. Do not re-run near-identical queries - they return the same results.
- Use the actual current date when a query needs a year. "latest X 2025" when it is 2026 retrieves stale answers that look fresh.
- Snippets are not sources. Search results are teasers; fetch the page before quoting, comparing, or deciding anything nontrivial from it.
- Conflicting or thin results -> search more, not "the sources disagree". Two more queries usually resolve what one left ambiguous.
Effort scaling
Single fact: 1 fetch. Comparison or medium question: 3-5. Genuine research: 5-10 with a short plan first (which sources, what order, what settles it). If it clearly needs far more, say so and propose the deeper process instead of thrashing.
Trust calibration
Treat fetched sources as EVIDENCE, not truth. Weigh three things before a claim graduates into your answer: authority (primary source > official docs > reputable secondary > random blog > forum hearsay), date (a current-looking page can carry stale content - check when it was written, not when it was served), and independence (three results parroting the same press release are one source, not three). Surprising claims from a single non-primary source get a second, independent confirmation before you repeat them. Raise extra skepticism in SEO-farmed spaces (product recommendations, "best X" listicles) and contested/conspiracy-prone topics. Never invent an attribution: if you are not sure a source says it, do not cite it as saying it.
After answering from search, the claims are still claims - verify load-bearing ones against the live system where one exists (see [[live-state-truth]], and [[jina-reader]] for fetching pages that block or garble a plain fetch).