feat: retrieve embeddings from database only when necessary

When performing a similarity search without using maximal marginal
relevance, the database query includes the embeddings by default,
whereas the retrived embeddings are discarded without use.

This can be very suboptimal when retrieve a large number of documents
due to communication overhead.

Author: fangyi-zhou

langchain_postgres/vectorstores.py

@@ -1060,9 +1060,9 @@ def _results_to_docs_and_scores(self, results: Any) -> List[Tuple[Document, floa
         docs = [
             (
                 Document(
-                    id=str(result.EmbeddingStore.id),
-                    page_content=result.EmbeddingStore.document,
-                    metadata=result.EmbeddingStore.cmetadata,
+                    id=str(result.id),
+                    page_content=result.document,
+                    metadata=result.cmetadata,
                 ),
                 result.distance if self.embeddings is not None else None,
             )
@@ -1395,8 +1395,16 @@ def __query_collection(
         embedding: List[float],
         k: int = 4,
         filter: Optional[Dict[str, str]] = None,
+        retrieve_embeddings: bool = False,
     ) -> Sequence[Any]:
         """Query the collection."""
+        columns_to_select = [
+            self.EmbeddingStore.id,
+            self.EmbeddingStore.document,
+            self.EmbeddingStore.cmetadata,
+        ]
+        if retrieve_embeddings:
+            columns_to_select.append(self.EmbeddingStore.embedding)
         with self._make_sync_session() as session:  # type: ignore[arg-type]
             collection = self.get_collection(session)
             if not collection:
@@ -1417,7 +1425,7 @@ def __query_collection(
 
             results: List[Any] = (
                 session.query(
-                    self.EmbeddingStore,
+                    *columns_to_select,
                     self.distance_strategy(embedding).label("distance"),
                 )
                 .filter(*filter_by)
@@ -1438,8 +1446,16 @@ async def __aquery_collection(
         embedding: List[float],
         k: int = 4,
         filter: Optional[Dict[str, str]] = None,
+        retrieve_embeddings: bool = False,
     ) -> Sequence[Any]:
         """Query the collection."""
+        columns_to_select = [
+            self.EmbeddingStore.id,
+            self.EmbeddingStore.document,
+            self.EmbeddingStore.cmetadata,
+        ]
+        if retrieve_embeddings:
+            columns_to_select.append(self.EmbeddingStore.embedding)
         async with self._make_async_session() as session:  # type: ignore[arg-type]
             collection = await self.aget_collection(session)
             if not collection:
@@ -1460,7 +1476,7 @@ async def __aquery_collection(
 
             stmt = (
                 select(
-                    self.EmbeddingStore,
+                    *columns_to_select,
                     self.distance_strategy(embedding).label("distance"),
                 )
                 .filter(*filter_by)
@@ -1899,9 +1915,11 @@ def max_marginal_relevance_search_with_score_by_vector(
                 relevance to the query and score for each.
         """
         assert not self._async_engine, "This method must be called without async_mode"
-        results = self.__query_collection(embedding=embedding, k=fetch_k, filter=filter)
+        results = self.__query_collection(
+            embedding=embedding, k=fetch_k, filter=filter, retrieve_embeddings=True
+        )
 
-        embedding_list = [result.EmbeddingStore.embedding for result in results]
+        embedding_list = [result.embedding for result in results]
 
         mmr_selected = maximal_marginal_relevance(
             np.array(embedding, dtype=np.float32),
@@ -1947,10 +1965,14 @@ async def amax_marginal_relevance_search_with_score_by_vector(
         await self.__apost_init__()  # Lazy async init
         async with self._make_async_session() as session:
             results = await self.__aquery_collection(
-                session=session, embedding=embedding, k=fetch_k, filter=filter
+                session=session,
+                embedding=embedding,
+                k=fetch_k,
+                filter=filter,
+                retrieve_embeddings=True,
             )
 
-            embedding_list = [result.EmbeddingStore.embedding for result in results]
+            embedding_list = [result.embedding for result in results]
 
             mmr_selected = maximal_marginal_relevance(
                 np.array(embedding, dtype=np.float32),