What it is
The semantic index is a vector-enriched layer over Microsoft 365 content — built ON TOP of the classic search index, adding meaning-based retrieval so "what did we agree about pricing" finds the doc that never says 'agree' or 'pricing'. Tenant-level (shared corpus, permission-trimmed at QUERY time) plus user-level slices (your mail/chats). It builds automatically for Copilot-licensed tenants; you don't size it, shard it or pay for it separately — you only live with its behaviours:
The behaviours that generate tickets
| Mystery | Mechanism |
|---|---|
| "I created the doc an hour ago, Copilot can't see it" | Index refresh lag — new content flows in on the indexing pipeline's cadence (minutes-to-hours by artifact type), not instantly |
| "I was JUST granted access, still nothing" | Permission changes propagate to query-time trimming on their own schedule — access granted ≠ immediately retrievable |
| "It cited a deleted file" | Index entries outlive deletions briefly; the citation link 404s while the snippet lingers — the ghost window |
| "It found the doc but summarised it badly" | Retrieval got the file; extraction quality varies by format (clean Word text beats a 40MB scanned PDF every time) |
| "Archived/records content doesn't appear" | Some storage states and content types sit outside grounding scope — verify per type, don't assume everything M365 is groundable |
Design consequences
- Content hygiene is retrieval engineering now: titles, headings and real text (not screenshots of text) directly improve Copilot answers — the first genuinely new reason to care about document quality since search.
- Restricted Content Discovery (governance module) works at THIS layer — sites excluded from grounding stay accessible but unfindable to Copilot; the index is where that scalpel cuts.
- Don't build SLAs on index freshness — no admin lever forces a reindex of specific content; automation that assumes read-after-write over Copilot will flake.
What to watch (proofs)
- Your tenant's real lag, measured: create → ask → ask again later (the foundations concept's test, formalised per artifact type: file, mail, chat). Write the numbers down; they end 'it's broken' debates.
- The permission-propagation lag: grant access → time until retrievable — the second number worth knowing.
- The ghost window: delete a test doc, ask about it — how long until the citation dies in YOUR tenant.
- Extraction quality: the same content as clean docx vs scanned PDF, same prompt — the demo that sells content hygiene to comms teams.