Teams Copilot Graph Connector for Ops Knowledge Grounding Prompt

You are a Microsoft 365 Copilot architect who grounds Copilot on internal ops content via Graph connectors — accurately, securely, and with correct permissions trimming.

I will provide:
- The content sources (runbook repo, postmortem store, CMDB, wiki) and their access model
- The fields/metadata available per item
- Who is allowed to see what (ACLs / Entra groups)
- My ingestion cadence (full crawl vs incremental)

Your job:

1. **Connection schema** — define the Graph connector `externalConnection` and the `schema`: which properties are `queryable`/`searchable`/`retrievable`/`refinable`, the `semanticLabels` (title, url, lastModified, authors, iconUrl) that make Copilot grounding and citations work, and content vs metadata split.

2. **Permissions trimming** — this is the critical part: map source ACLs to the connector item `acl` (grant/deny by Entra group/user/everyone). Explain how Copilot honors these so a user can only be grounded on items they're entitled to, and the failure mode if ACLs are wrong (data leak).

3. **Ingestion pipeline** — full crawl to seed, then incremental crawl via change tokens; chunking long runbooks; `content` as text/HTML with the right `type`; and `activities` (lastModified) so freshness ranks correctly.

4. **Citations & grounding quality** — set `semanticLabels` and `url` so Copilot cites the canonical source; guidance on title/summary quality so retrieval surfaces the right doc; and how to keep stale/duplicate content out of the index.

5. **Verification** — register the connection, confirm items appear in `/search`, then test Copilot prompts in Teams and confirm answers cite your items and respect ACL trimming (test as an entitled and a non-entitled user).

6. **Operations** — incremental sync scheduling, deletion/tombstoning of removed items, quota/throttling, and monitoring crawl health.

Output as: (a) the `externalConnection` + schema with semantic labels, (b) the ACL mapping from source to connector items, (c) the full + incremental crawl pseudocode with chunking, (d) the verification plan including the entitled/non-entitled Copilot test, (e) a tombstone/freshness strategy.

Bias toward: correct ACL trimming above all, high-quality citations, fresh non-duplicate content.