feat(vector): add direct pgvector adapter

This commit is contained in:
2026-07-12 01:01:15 +02:00
parent ebdd3aa2c5
commit b09341f07e
7 changed files with 515 additions and 35 deletions
@@ -0,0 +1,50 @@
# Local pgvector Task 1 report
## Status
Implemented the direct `PgVectorStore` behind the transport-neutral `VectorStore` port.
The adapter uses separate optional reader and writer database configurations, derives
capabilities from configured authority, validates strict positive search limits, filters kinds
in SQL before limiting, and merges multi-collection results by cosine similarity.
All collection identifiers are selected from the fixed `schema_records`, `evidence`, and
`memory` allowlist and composed with `psycopg2.sql.Identifier`. Values, vectors, kinds, hashes,
and limits remain bound parameters. Collection/kind mismatches fail with `VectorStoreError`.
Upserts preserve the canonical metadata shape, use `record_key` conflict semantics, update the
transport hash and embedding, and leave semantic metadata fields intact. Health probes reader
and writer independently and reports observed `vector(N)` dimensions against the configured
embedding dimension.
## Configuration and factory
`pgvector_direct` now accepts explicit optional `reader` and `writer` `DatabaseConfig` entries.
The former `connection` entry remains supported as a deprecated read-only compatibility path.
`build_vector_store(..., require_write=True)` accepts writer-only direct configurations and
fails early when no explicit writer is present.
The transitional `build_vector_loader` bulk-sync path remains in place. It uses an explicit
direct writer when present, or the legacy `connection`; it deliberately does not treat a new
reader-only credential as writable. No production schema migration was added.
## TDD and verification
- RED: the new tests initially failed at collection because `PgVectorStore` did not exist.
- Docker L0 pgvector tests: `11 passed`.
- Direct + HTTP parity/factory/config focus: `51 passed`.
- Full harness: `461 passed, 5 deselected`.
- Changed-file Ruff lint: clean.
- Changed-file Ruff format check: clean.
- `git diff --check`: clean.
The repository-wide `ruff check .` still reports 34 pre-existing test-file findings outside
Task 1; none are in changed files. The full pytest suite emits 17 existing legacy-config
deprecation warnings.
## Scope and concerns
- Test fixtures create only the three existing vector tables needed to exercise the adapter;
migration/versioning remains Task 2.
- The legacy single `connection` form stays read-only through the public port, matching its
previous adapter behavior, while remaining available to the explicitly documented bulk-loader
transition.