Embedding Pipeline Service  ·  View 15 of 22  ·  Runtime

Retrieval Paths

Three consumer shapes and two degradation states across the same five stages — including the one failure that withholds results.

Editable source SVG draw.io All views
Resolve Embed query Candidate search Filter Return Workspace search Alias -> index Same contract as index Vector top-200 Lexical top-200 ACL filter Suppression Top-50 + staleness Ask-your-docs Alias -> index Query embedded Vector top-200 ACL filter Top-10 + offsets Similar documents Alias -> index Reuse stored vector Neighbours, self excluded ACL filter Ranked documents Vector tier degraded Alias still resolves Skipped Lexical only ACL filter unchanged degraded = true ACL authority down Alias resolves Query embedded Candidates found Fail closed withheld = N Retrieval Paths by Consumer and Degradation State Four degradations are available before an error: no re-rank, no vectors, no cache, sampled telemetry. Only a permission failure withholds results. v 1.0 · owner Data & AI Platform Architecture · date 2026-10

Degrade a dimension, not the service

  • Four degradations are available before an error: skip re-ranking, drop to lexical-only, lose the query cache, sample telemetry. Each is visible in the response.
  • A vector-tier outage returns lexical results with degraded = true rather than a 503. A consumer that wants to hide those results can; the platform does not decide for it.
  • Only a permission-resolution failure withholds: results are excluded and the count of withheld items is returned.

Why similar-documents is cheaper

  • It reuses the document's already-stored vector instead of embedding a query, which removes the one GPU call on the read path.
  • That is only correct because the stored vector and the index share a contract — the same invariant that forbids a query spanning contracts.

Open question carried forward

  • Where the permission filter belongs (requirement question 5) is drawn here as post-filter over a larger candidate set. With a user who can read 1% of a tenant's corpus, recall loss may force pre-filtering or per-principal partitions.
  • Candidate set sizes (200 → 50) are a starting point to be measured per corpus, not a settled constant.
◀  Embedding Pipeline All views Critical Flow — Contract Migration and Alias Cutover  ▶
Solution Architecture v1.0 · open source, self-hosted on Kubernetes · Data & AI Platform Architecture · 2026-10 · 22 views · 17 architecture decision records15-retrieval-paths