Image vector embedding

app/api/routers/user_products.py

@@ -477,13 +477,15 @@ def _brand_defaults(brand_parent: str) -> Dict[str, Any]:
     """
     # Two reads on purpose, and the split matters on a memory-capped host.
     #
-    # `_brand_sample` is SELECT * limit 1 - one row, embedding and all, because
-    # the fields it seeds (category, price band, size) need the whole row.
+    # `_brand_sample` is one full row (limit 1), because the fields it seeds
+    # (category, price band, size) need the whole row. The product readers now
+    # project the vector columns out, so "full" no longer means "with the
+    # embedding".
     #
     # The consensus read is 300 rows, so it takes only the two columns it
-    # actually inspects. Measured: SELECT * over 244 rows costs 3.0 MB, of
-    # which 4.7 KB per row is an embedding string nothing here reads. Two named
-    # columns is roughly 50 KB for the same rows.
+    # actually inspects. Measured when reads were still SELECT *: 244 rows
+    # cost 3.0 MB, of which 4.7 KB per row was an embedding string nothing
+    # here reads. Two named columns is roughly 50 KB for the same rows.
     rows = consensus_rows(brand_parent, ["fssai_license", "providers"],
                           limit=_CONSENSUS_SAMPLE_LIMIT)