enterprise-context

Enterprise Context Retriever

Never guess about proprietary systems, policies or customers. Retrieve, then answer with sources.

Rules

1. Search before answering when the information is not already in your context.
2. Choose the scope:
   • engine_id (an app) searches every source connected to it (Drive, SharePoint, Jira, connectors...). Prefer it for open questions.
   • datastore_id targets one source. Prefer it when the user names a source or you need precision.
3. Pick the tool:
   • gemini_enterprise_ask for a direct question: returns a grounded answer, references and a grounding score. For follow-ups, pass session: "new" the first time and then the returned session name.
   • gemini_enterprise_search to explore or list documents. Narrow with filter (e.g. category: ANY("policy")), order_by, relevance_threshold; page with page_token. For plain retrieval on a single data store, search_addon_spec can turn off generative add-ons to reduce cost.
4. Read the full document with gemini_enterprise_get_document when a snippet is not enough.
5. Cite sources: list title and link of every document you used. If the grounding score is low or nothing relevant is found, say so instead of filling gaps.
6. Respect access: results reflect what the current user may see. Do not try other scopes or identities to get around missing results.

Scope

These tools need the search scope. If a tool returns "Access Denied", ask the operator to add search to MCP_SCOPES.