Grounded Search

/enterprise-search

Answer one question from the material the user provides -- attached files, pasted documents, exported chat threads -- plus web search when current external facts matter. Decompose the question, search every source, and synthesize a single sourced answer.

Lane boundaries: this is targeted question-answering over a provided corpus. For an open-ended, multi-step research deliverable, use deep-research. For designing a search plan across connected workplace tools, use search-strategy.

Instructions

1. Inventory the corpus

List what the user has provided: each attachment, pasted block, or linked excerpt, with a one-line note on what it appears to be (spec, thread export, notes, contract). If the question obviously needs a source they have not provided ("what did legal say?" with no legal doc present), say so up front rather than answering from the wrong material.

2. Parse the question

• Intent: a decision, a fact/number, a document location, a person's statement, or an exploratory "what do we know about X".
• Entities: people, projects, features, dates mentioned.
• Time constraints: "latest", "before the launch", specific dates -- use document dates and internal timestamps to honor them.
• Scope: corpus-only, or corpus + web? Default to corpus-only; add web search when the question involves external or current facts (versions, prices, news) or the user asks.

3. Search every source

For each provided document, scan for the entities and their synonyms (a "refund policy" may appear as "money-back", "cancellation terms"). Do not stop at the first hit; the corpus is small enough to be exhaustive. For web supplementation, run targeted queries and capture publisher + date.

4. Rank and deduplicate

• Group the same fact appearing in multiple sources; prefer the most authoritative and most recent version.
• Authority within a corpus: signed/final docs > drafts > meeting notes > chat messages for factual questions; conversations outrank docs for "what did we discuss" questions.
• Note conflicts explicitly ("the spec says 100 rps; the thread from a week later says 250").

5. Present a synthesized answer

Factual/decision queries:

[Direct answer]

Sources:
- [doc/attachment name, section or line, date if known]
- [web source, publisher, date]

Exploratory queries:

[Synthesized summary]

Found across:
- [doc 1]: [what it contributed]
- [doc 2]: [what it contributed]
- Web: [what it contributed]

"Find" queries: the exact passage, quoted, with its document and location, plus related passages found elsewhere.

6. Edge cases

• Ambiguous question: ask one clarifying question before searching ("'API redesign' -- the REST v2 doc or the SDK notes?").
• No results: say what you searched and suggest broader terms or the missing source type.
• Answer depends on absent material: answer from what exists, flag the gap, and name the document that would close it.

Rules

• Every claim in the answer traces to a named source; never blend corpus facts with unsourced recall.
• Quote exactly when the wording matters (policies, commitments, numbers).
• Do not summarize the whole corpus unasked -- answer the question.

Usage

/enterprise-search $ARGUMENTS