perf(search): pg_trgm fuzzy staff search — typo-tolerant, multi-word, ~100ms (was ILIKE seq scan)

The Postgres answer to WowPlatter's sluggish rigid search. Was ILIKE '%..%' → Seq Scan (3038 cost),
zero typo tolerance, and multi-word phrases ('bob marley legend 2022') returned NOTHING (needs a
contiguous substring). Now: pg_trgm GIN index on disc_release.search_text (already populated) +
word_similarity operators — :q <% search_text (filter, GIN-accelerated) ranked by :q <<-> search_text.
Threshold pinned at 0.45 on the DB (recall vs speed sweet spot ~140ms; 0.3 too loose/slow, 0.6 too strict).
Added the missing disc_release_format(release_id) index (was seq-scanning per row) + inventory.title trgm
for stock-finder. All DDL idempotent in _STARTUP_DDL. Live: 'bob marly survivl' (2 typos) -> Survival in 117ms.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/intake_routes.py

@@ -184,10 +184,13 @@ async def stage(body: StageIn, ident=Depends(require_token), db=Depends(get_db))
 async def search(q: str = Query(""), label: str = Query(""), year: int | None = None,
                  country: str = Query(""), fmt: str = Query(""),
                  ident=Depends(require_token), db=Depends(get_db)):
-    """Search the local Discogs mirror — the fast daily path (re-stocks resolve instantly)."""
+    """Search the local Discogs mirror — fast + typo-tolerant via pg_trgm on search_text
-    where, params = ["TRUE"], {}
+    (word_similarity: multi-word, any order, fuzzy; GIN-indexed; threshold 0.3 pinned on the DB)."""
+    where, params, order = ["TRUE"], {}, "dr.title"
     if q:
-        where.append("(dr.title ILIKE :q OR dr.artists_sort ILIKE :q)"); params["q"] = f"%{q}%"
+        where.append(":q <% dr.search_text")          # typo-tolerant word match, GIN-accelerated
+        order = ":q <<-> dr.search_text"               # rank by word distance (closest first)
+        params["q"] = q
     if year:
         where.append("dr.year = :y"); params["y"] = year
     if country:
@@ -207,7 +210,7 @@ async def search(q: str = Query(""), label: str = Query(""), year: int | None =
                EXISTS (SELECT 1 FROM inventory i WHERE i.release_id=dr.id AND i.in_stock AND i.store_id=1) AS in_stock
         FROM disc_release dr LEFT JOIN disc_cache dc ON dc.release_id=dr.id
         WHERE {' AND '.join(where)}
-        ORDER BY dr.title LIMIT 40"""), params)).mappings()]
+        ORDER BY {order} LIMIT 40"""), params)).mappings()]
     return {"items": rows, "source": "local"}
 
 

app/main.py

@@ -91,6 +91,13 @@ _STARTUP_DDL = [
     "CREATE SEQUENCE IF NOT EXISTS sale_items_id_seq",
     "ALTER TABLE sale_items ALTER COLUMN id SET DEFAULT nextval('sale_items_id_seq')",
     "SELECT setval('sale_items_id_seq', GREATEST(1, (SELECT coalesce(max(id),0) FROM sale_items)))",
+    # fast fuzzy staff search — pg_trgm GIN indexes (typo-tolerant, multi-word, index-accelerated;
+    # replaces the ILIKE '%..%' seq scans). word_similarity threshold pinned on the DB (0.3 = good recall).
+    "CREATE EXTENSION IF NOT EXISTS pg_trgm",
+    "ALTER DATABASE recordgod SET pg_trgm.word_similarity_threshold = 0.45",  # recall vs speed sweet spot (~140ms)
+    "CREATE INDEX IF NOT EXISTS disc_release_search_trgm ON disc_release USING gin (search_text gin_trgm_ops)",
+    "CREATE INDEX IF NOT EXISTS inventory_title_trgm ON inventory USING gin (title gin_trgm_ops)",
+    "CREATE INDEX IF NOT EXISTS disc_release_format_release_id_idx ON disc_release_format(release_id)",
 ]