joaoh82 / rust_sqlite / 25040177500

69%

Build Type

push

github

Committed by web-flow

Commit Message

Phase 7d.2: SQL integration for HNSW (CREATE INDEX, KNN probe) (#50)

Wires the Phase 7d.1 HNSW algorithm into the SQL surface so users
can finally do KNN queries through the engine:

  CREATE TABLE docs (id INTEGER PRIMARY KEY, e VECTOR(384));
  INSERT INTO docs (e) VALUES ([0.1, 0.2, ...]);
  CREATE INDEX ix_e ON docs USING hnsw (e);
  SELECT id FROM docs
  ORDER BY vec_distance_l2(e, [0.5, 0.3, ...])
  LIMIT 10;

Five integration points:

  1. **Table::hnsw_indexes** — new Vec<HnswIndexEntry> field
     parallel to secondary_indexes. HnswIndexEntry holds
     {name, column_name, index} where index is the in-memory
     HnswIndex from 7d.1. Auto-cloned by deep_clone (Phase 4f
     transaction snapshot path).

  2. **CREATE INDEX … USING hnsw (col)** parser + executor
     branch. Routes through new IndexMethod::{Btree, Hnsw}
     dispatch in execute_create_index. Validates the column is
     VECTOR(N) and rejects UNIQUE on HNSW (no semantic meaning).
     Pre-populates the graph from existing rows. Distance metric
     is L2 by default; cosine and dot are 7d.x follow-ups
     (would need either WITH (metric=…) clause or hnsw_cosine
     / hnsw_dot named methods — not in 7d.2 scope).

  3. **INSERT incremental maintenance** in Table::insert_row.
     After writing a row, calls Table::maintain_hnsw_on_insert
     for any HNSW index whose column matches. Uses a snapshot
     of current row storage as the get_vec source so the
     algorithm's neighbor lookups don't fight with the
     &mut self.hnsw_indexes borrow.

  4. **DELETE / UPDATE refusal** in execute_delete /
     execute_update. HNSW lacks an in-place delete-node
     primitive and we don't yet have a graph-rebuild trigger
     (that's part of 7d.3 with persistence). Fail with a
     helpful message that points at the workaround
     (DROP the index, mutate, re-CREATE) and explicitly
     mentions Phase 7d.3 as the follow-up.

  5. **Query optimizer hook** in execute_select_rows. Before
     ... (continued)

Coverage Stats

218 of 251 new or added lines in 4 files covered. (86.85%)

3 existing lines in 1 file now uncovered.

4973 of 7076 relevant lines covered (70.28%)

1.45 hits per line