Embedding Pipeline Service · View 04 of 22 · People and journeys
The trough, and what answers it
- The dip is at phase 4: the result is correct and the reader cannot interpret it. A snippet that is one paragraph with no heading above it looks like the wrong document.
- Both answers are structural and both are decided before any retrieval happens — offsets preserved at extraction, and the heading path carried on the chunk.
- The second pain, old wording still matching, is the atomic-visibility requirement: a half-reindexed document is worse than a stale one.
What the platform promises here
- Interactive lane: the edit is retrievable at p50 30 s, p95 2 min, p99 10 min (stated assumption).
- Retrieval p50 45 ms, p99 180 ms in-region for a hybrid top-50 over a tenant partition.
- Every response carries the document version and a staleness indication, so the surface can tell the reader when it is behind.
What it does not promise
- Nothing in the editor tells the user that indexing is in progress. That is a product gap this architecture enables an answer to and does not supply.
- Phase 2 assumes the user does not watch. If a product adds a "search as you type in the same session" surface, the freshness budget it needs is a different conversation.