Docs follow-ups from the Milvus tire-kick session (#1235)
Verified gaps in docs/semantic-search.md (and adjacent), all observed hands-on:
Positive findings worth stating in docs while editing: identical vector rankings and
scores (4-decimal agreement) across sqlite-vec/pgvector/Milvus Lite on a 222-entity
corpus; dimension-mismatch refusal message is excellent; missing-extra error is excellent;
30s timeout honored on unreachable servers; collections reload across processes (#1185).
Docs follow-ups from the Milvus tire-kick session (#1235)
Verified gaps in
docs/semantic-search.md(and adjacent), all observed hands-on:test infrastructure; a user enabling Milvus must reverse-engineer
docker-compose-postgres.yml+BASIC_MEMORY_DATABASE_BACKEND/DATABASE_URLfrom thejustfile. One short "run Postgres locally" section would unblock the whole Milvus path.
semantic_vector_indexrequiresbm reindex --embeddings(incremental suffices — verified: 8,689 chunks re-owned in~3 min); the docs only mention reindex in the migration-cleanup narrative.
nothing and hybrid quietly serves FTS-only until the reindex runs.
.dbreality: pymilvus hard-requires the.dbsuffix and an existingparent directory, and what it creates is a directory tree (collections/, databases/,
LOCK), not a file. Errors for bad shapes are raw pymilvus exceptions with no pointer to
the
milvus_urisetting.milvus_databaseis ignored by Milvus Lite (no named-database support); it'spassed unconditionally as
db_name.(
search_vector_embeddings, 8,689 rows in the session) remains on disk indefinitely;document the manual cleanup or note the disk cost until a cleanup command exists.
Positive findings worth stating in docs while editing: identical vector rankings and
scores (4-decimal agreement) across sqlite-vec/pgvector/Milvus Lite on a 222-entity
corpus; dimension-mismatch refusal message is excellent; missing-extra error is excellent;
30s timeout honored on unreachable servers; collections reload across processes (#1185).