feat: add vertex vector search datapoints deletion with metadata filt…

…ers support (#559)

libs/vertexai/langchain_google_vertexai/vectorstores/_searcher.py

@@ -20,6 +20,8 @@
     to_data_points,
 )
 
+MAX_DATA_POINTS = 10000
+
 
 class Searcher(ABC):
     """Abstract implementation of a similarity searcher."""
@@ -60,6 +62,22 @@ def add_to_index(
         """
         raise NotImplementedError()
 
+    @abstractmethod
+    def remove_datapoints(
+        self,
+        datapoint_ids: List[str],
+        **kwargs: Any,
+    ) -> None:
+        raise NotImplementedError()
+
+    @abstractmethod
+    def get_datapoints_by_filter(
+        self,
+        metadata: dict,
+        max_datapoints: int = MAX_DATA_POINTS,
+    ) -> List[str]:
+        raise NotImplementedError()
+
     def _postprocess_response(
         self, response: List[List[MatchNeighbor]]
     ) -> List[List[Tuple[str, float]]]:
@@ -105,6 +123,31 @@ def __init__(
         self._staging_bucket = staging_bucket
         self._stream_update = stream_update
 
+    def get_datapoints_by_filter(
+        self,
+        metadata: dict,
+        max_datapoints: int = MAX_DATA_POINTS,
+    ) -> List[str]:
+        """Gets all the datapoints matching the metadata filters (text only)
+        on the specified deployed index.
+        """
+        index_config = self._index.to_dict()["metadata"]["config"]
+        embeddings = [[0.0] * int(index_config.get("dimensions", 1))]
+        filter_ = [
+            Namespace(name=key, allow_tokens=[value]) for key, value in metadata.items()
+        ]
+        neighbors = self.find_neighbors(
+            embeddings=embeddings, k=max_datapoints, filter_=filter_
+        )
+        return [_id for (_id, _) in neighbors[0]] if neighbors else []
+
+    def remove_datapoints(
+        self,
+        datapoint_ids: List[str],
+        **kwargs: Any,
+    ) -> None:
+        self._index.remove_datapoints(datapoint_ids=datapoint_ids)
+
     def add_to_index(
         self,
         ids: List[str],

libs/vertexai/langchain_google_vertexai/vectorstores/vectorstores.py

@@ -119,6 +119,8 @@ def similarity_search_by_vector_with_score(
         neighbors_list = self._searcher.find_neighbors(
             embeddings=[embedding], k=k, filter_=filter, numeric_filter=numeric_filter
         )
+        if not neighbors_list:
+            return []
 
         keys = [key for key, _ in neighbors_list[0]]
         distances = [distance for _, distance in neighbors_list[0]]
@@ -134,6 +136,36 @@ def similarity_search_by_vector_with_score(
             message = f"Documents with ids: {missing_docs} not found in the storage"
             raise ValueError(message)
 
+    def delete(self, ids: Optional[List[str]] = None, **kwargs: Any) -> Optional[bool]:
+        """
+        Delete by vector ID.
+        Args:
+            ids (Optional[List[str]]): List of ids to delete.
+            **kwargs (Any): If added metadata={}, deletes the documents
+            that match the metadata filter and the parameter ids is not needed.
+        Returns:
+            Optional[bool]: True if deletion is successful.
+        Raises:
+            ValueError: If ids is None or an empty list.
+            RuntimeError: If an error occurs during the deletion process.
+        """
+        metadata = kwargs.get("metadata")
+        if (not ids and not metadata) or (ids and metadata):
+            raise ValueError(
+                "You should provide ids (as list of id's) or a metadata"
+                "filter for deleting documents."
+            )
+        if metadata:
+            ids = self._searcher.get_datapoints_by_filter(metadata=metadata)
+            if not ids:
+                return False
+        try:
+            self._searcher.remove_datapoints(datapoint_ids=ids)  # type: ignore[arg-type]
+            self._document_storage.mdelete(ids)  # type: ignore[arg-type]
+            return True
+        except Exception as e:
+            raise RuntimeError(f"Error during deletion: {str(e)}") from e
+
     def similarity_search(
         self,
         query: str,