[PGVector] Unnecessary columns queried leads to high latency

[JoshuaPostel]

langchain-postgres/langchain_postgres/vectorstores.py

Lines 1419 to 1432 in 18b1bcd

	results: List[Any] = (
	session.query(
	self.EmbeddingStore,
	self.distance_strategy(embedding).label("distance"),
	)
	.filter(*filter_by)
	.order_by(sqlalchemy.asc("distance"))
	.join(
	self.CollectionStore,
	self.EmbeddingStore.collection_id == self.CollectionStore.uuid,
	)
	.limit(k)
	.all()
	)

This query returns the embedding column which is expensive to deserialize into a python object and AFAICT is not used downstream. From my observations benchmarking against an equivalent query without the embedding column, this costs ~150ms (1024 dimension vector, k=30).

[averikitsch]

The new PGVectorStore interface includes this improvement.