Tracking issue for the manual end-to-end verification session of the Milvus vector-index
adapter (SPEC-81; PRs #1141, #1158, #1185). There is no automated end-to-end coverage —
every Milvus test uses fakes or bare Milvus Lite — so this session exercises the real
stack: Postgres + Milvus Lite through sync → manifest → search, adapter switches, reindex,
and failure drills, using a realistic 200+ note corpus.
Session checklist
Deliverables
Tracking issue for the manual end-to-end verification session of the Milvus vector-index
adapter (SPEC-81; PRs #1141, #1158, #1185). There is no automated end-to-end coverage —
every Milvus test uses fakes or bare Milvus Lite — so this session exercises the real
stack: Postgres + Milvus Lite through sync → manifest → search, adapter switches, reindex,
and failure drills, using a realistic 200+ note corpus.
Session checklist
pgvector / Milvus Lite; compare top-5 membership, cosine scores, hybrid rankings,
latency (~10 known-answer queries)
bm reindex --embeddings; collection naming verified0, hybrid silently degrades to FTS (documented UX finding)
vector_indexownership flip; stale pgvector rows accountingSemanticVectorIndexExtensionErrorcapture + recovery UXtime, timeout/lock-hold), dimension mismatch refusal
milvus_token/milvus_uriin config output)Deliverables
.dbURI requirements, Lite
milvus_databasecaveat, degraded-window behavior)registry removed, namespace = DB-locator hash, collection identity excludes model/dims)