es-search

Fan one question out across every connected source and return a single ranked, deduplicated, attributed result set.

SKILL.md

GitHub

Enterprise Search

When to use

Reach for this when someone asks "where is..." or "what did we say about..." and the answer could live in any connected source — chat, email, cloud storage, or issue trackers. You run one fan-out and hand back one answer instead of making the user hop tools.

Not for: planning a fuzzy, multi-part question (use es-search-strategy first); merging conflicting hits into a narrative (use es-knowledge-synthesis); periodic activity roundups (use es-digest); or auditing what is even connected (use es-source-management).

Method

  1. Read the intent. Extract the core entities, the likely artifact type (message, file, thread, ticket), and any person or time constraint stated or implied.
  2. Enumerate live sources. List which connected sources are up. Decision: if a source structurally cannot hold this artifact type, skip it and say so — do not burn a query on it.
  3. Fan out in parallel. Issue one tailored query per relevant source. Use keyword operators for storage and trackers, semantic phrasing for chat and email. Decision: if a source returns zero, loosen its terms once before dropping it.
  4. Normalize hits to a common shape: title, source, author, timestamp, snippet, link.
  5. Rank globally by term match, recency, source/author authority, and how directly the snippet answers the question — never by which source replied first.
  6. Deduplicate. Fold the same artifact surfaced by multiple sources into one entry that notes every place it appears.

Example

Ask: "Where's the final Q3 pricing deck?" You fan out: storage (keyword Q3 pricing deck filetype:deck), chat (semantic "final Q3 pricing"), email ("Q3 pricing attachment"). Storage returns two versions and Slack a link to one of them. You dedupe the Slack link into the storage hit, rank the file titled "Q3-Pricing-FINAL" (yesterday) above "Q3-Pricing-draft" (three weeks old), and answer with the FINAL link plus the draft as a runner-up.

Pitfalls

  • Reusing one query string verbatim across sources — keyword syntax starves semantic sources and vice versa.
  • Ranking by source order or first-responder speed instead of relevance and recency.
  • Reporting the first hit as the answer when two near-identical artifacts exist; surface both.
  • Silently skipping an unconnected source, leaving the user to assume it was searched.

Output format

Answer: <one sentence if results settle it, else "Top candidates below">

Ranked results:
1. [source] Title — Author, YYYY-MM-DD
   "snippet"  <link>
2. ...

Coverage: searched <sources>; skipped <source> (<reason>)
Refine: <2-3 narrowing filters — person / date / source>