Search Index Capacity

Azure AI Search is already provisioned and in production on the Free SKU (prod-srch-esposter-001, with a dev twin), serving esbabbler message search and the Sent tab through the single messages-index — see Azure services. The deferred item is not the service; it is paying for capacity once the Free SKU's ceiling is reached.

The Free SKU allows 3 indexes, 3 indexers, 3 data sources, and 50 MB of storage in total, on shared infrastructure with no SLA and no scale-out — replicas and partitions are not configurable on this tier — and caps a single indexing invocation at 10,000 documents. Storage is the ceiling that matters — there is no separate per-index document quota to reach first. Today one index and one indexer are in use, so two slots are spare and the storage cap is the binding constraint — message documents carry the body text, attachment filenames, and sender name, so index growth tracks message volume directly.

What happens at the ceiling

Nothing in the app breaks loudly. The messages-indexer is a scheduled pull indexer, so when it can no longer write documents its runs start failing while every already-indexed document stays queryable: search keeps working and silently stops including new messages. The failure is visible only in the indexer's status endpoint (GET /indexers/messages-indexer/status), which is exactly the surface Observability does not alert on. Assume the first symptom is a user reporting that a recent message can't be found.

Why deferred

• Message volume is nowhere near the storage cap, and the two spare index slots absorb the next indexing idea without any change.
• The tier of an Azure AI Search service cannot be changed after creation — there is no Free → Basic upgrade. Moving up means standing up a second service, re-creating the index and indexer from packages/infra/data/searchIndexes/messages-index.json, letting it backfill, and repointing the app. That is a real migration, not a SKU edit, and it is not worth rehearsing before the numbers demand it.
• Basic is the first paid tier and costs tens of dollars a month — an order of magnitude above every other line item in this estate, whose whole posture is the $0.01 budget guard.

Revisit when

The indexer's status endpoint reports a quota or storage failure, index storage passes roughly half the 50 MB cap, or a second feature needs an index and would take the last free slot (the pending candidate is resource search).

Cheaper interim

Bounding the corpus buys capacity before spend does: message retention pruning old rows shrinks what the indexer feeds, and keeping resource search on pg_trgm (global search relevance) leaves the two spare index slots unspent.