fix(index): clear the legacy embeddings table on reset and deindex

reset()'s docstring warns "leaving stale rows here means semantic
search can return orphaned hits after a reindex" and deindex()'s
docstring promises to remove "the embedding row for any kind," but
neither ever touched the legacy embeddings table alongside
embedding_index. every artifact deletion or full reindex leaked a
stale vector row that never gets cleaned up.

search_embeddings (plural, the reader of the legacy table) currently
has no callers in src/, so this isn't live search pollution today -
but health._check_orphan_embeddings explicitly checks both tables, so
a leaked row permanently trips the orphan_embedding fsck warning with
no way to clear it via the documented remedy (reindexing), and grows
state.db unbounded over a kb's lifetime.

add the missing DELETE FROM embeddings to reset()'s executescript, and
the missing per-row delete to deindex(), matching the existing
embedding_index handling in both.

this exact defect and fix were previously submitted as #543
(fix(index): clear the legacy embeddings table on reset and deindex),
but that PR was closed unmerged for going stale against a fast-moving
test branch, not for anything wrong with the change; the maintainer's
closing comment explicitly invited a fresh PR.

Fixes #687

Author: joaovictor91123

CHANGELOG.md

@@ -45,6 +45,13 @@ All notable changes to vouch are documented here. Format follows
   artifact the caller could not already retrieve, and it touches no write path.
 
 ### Fixed
+- **`reset()`/`deindex()` now clear the legacy `embeddings` table too**
+  (#543 reopened, root-caused): both functions' own docstrings promise to
+  remove every embedding row for a reindex or a deleted artifact, but
+  neither ever touched the legacy `embeddings` table alongside
+  `embedding_index` — a leaked row permanently tripped `fsck`'s
+  `orphan_embedding` warning with no way to clear it via reindexing, and
+  grew `state.db` unbounded over a KB's lifetime.
 - **`hub_client` ETag lookup is now case-insensitive** (#662): `_request`
   flattened `resp.headers` (case-insensitive by design) into a plain
   `dict`, so `pull()`'s `resp_headers.get("ETag")` silently returned

src/vouch/index_db.py

@@ -122,6 +122,7 @@ def reset(kb_dir: Path) -> None:
             "DELETE FROM claims_fts;"
             "DELETE FROM pages_fts;"
             "DELETE FROM entities_fts;"
+            "DELETE FROM embeddings;"
             "DELETE FROM embedding_index;"
             "DELETE FROM query_embedding_cache;"
             "DELETE FROM embedding_dupes;"
@@ -202,6 +203,9 @@ def deindex(conn: sqlite3.Connection, *, kind: str, id: str) -> None:
         conn.execute("DELETE FROM pages_fts WHERE id = ?", (id,))
     elif kind == "entity":
         conn.execute("DELETE FROM entities_fts WHERE id = ?", (id,))
+    conn.execute(
+        "DELETE FROM embeddings WHERE kind = ? AND id = ?", (kind, id)
+    )
     conn.execute(
         "DELETE FROM embedding_index WHERE kind = ? AND id = ?", (kind, id)
     )

tests/test_index_db_embeddings.py

@@ -210,6 +210,34 @@ def test_legacy_search_embeddings_ranks_by_cosine(store: KBStore) -> None:
     assert hits[0][1] == "c1"
 
 
+def test_deindex_removes_the_legacy_embeddings_row(store: KBStore) -> None:
+    # deindex()'s own docstring promises to remove "the embedding row for
+    # any kind" but only cleared embedding_index, never the legacy
+    # `embeddings` table search_embeddings (plural) reads — a deleted
+    # artifact's vector leaked there forever.
+    with index_db.open_db(store.kb_dir) as conn:
+        index_db.index_embedding(
+            conn, kind="claim", id="ghost", vec=_vec("gone now").tolist()
+        )
+        conn.commit()
+        index_db.deindex(conn, kind="claim", id="ghost")
+        conn.commit()
+    assert index_db.search_embeddings(store.kb_dir, _vec("gone now").tolist()) == []
+
+
+def test_reset_clears_the_legacy_embeddings_table(store: KBStore) -> None:
+    # reset()'s own docstring warns "leaving stale rows here means semantic
+    # search can return orphaned hits after a reindex" but never cleared
+    # the legacy `embeddings` table itself.
+    with index_db.open_db(store.kb_dir) as conn:
+        index_db.index_embedding(
+            conn, kind="claim", id="c1", vec=_vec("will be reset").tolist()
+        )
+        conn.commit()
+    index_db.reset(store.kb_dir)
+    assert index_db.search_embeddings(store.kb_dir, _vec("will be reset").tolist()) == []
+
+
 def test_legacy_search_embeddings_rejects_an_empty_query(store: KBStore) -> None:
     assert index_db.search_embeddings(store.kb_dir, []) == []