fix(memory): honor Vertex RAG top-k configuration

VertexAiRagMemoryService accepts a similarity_top_k argument but dropped it
when constructing the VertexRagStore, so search_memory read back None and
retrieval silently ignored the configured limit. The writer was removed
during the agentplatform migration while the reader was kept.

The value is now held on the service and passed to the RagQuery, which is
where retrieveContexts reads it. It is deliberately not restored on the
VertexRagStore: that request message carries rag_resources, rag_corpora and
vector_distance_threshold only, so a similarity_top_k set there would travel
as an undefined field on the request rather than as a retrieval limit.

Behavior change: anyone who set similarity_top_k has been getting the API
default and now gets the value they configured, so result counts move down or
up depending on whether that value is below or above the default.

No public interface changes.

Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 954790154

Author: GWeale

src/google/adk/memory/vertex_ai_rag_memory_service.py

@@ -136,6 +136,9 @@ def __init__(
         self._project = self._project or parts[1]
         self._location = self._location or parts[3]
 
+    # Top-k belongs on the retrieval query, not on the store: the
+    # retrieveContexts request's VertexRagStore has no such field.
+    self._similarity_top_k = similarity_top_k
     self._vertex_rag_store = types.VertexRagStore(
         rag_resources=[
             types.VertexRagStoreRagResource(rag_corpus=rag_corpus),
@@ -208,7 +211,7 @@ async def search_memory(
         vertex_rag_store=self._vertex_rag_store,
         query=agentplatform_types.RagQuery(
             text=query,
-            similarity_top_k=self._vertex_rag_store.similarity_top_k,
+            similarity_top_k=self._similarity_top_k,
         ),
     )
     memory_results = []

tests/unittests/memory/test_vertex_ai_rag_memory_service.py

@@ -22,6 +22,7 @@
 from google.adk.sessions.session import Session
 from google.genai import types
 import pytest
+from pytest_mock import MockerFixture
 
 
 def _rag_context(source_display_name: str, text: str) -> SimpleNamespace:
@@ -31,6 +32,34 @@ def _rag_context(source_display_name: str, text: str) -> SimpleNamespace:
   )
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("configured_top_k", [7, None])
+async def test_search_memory_forwards_similarity_top_k(
+    mocker: MockerFixture,
+    configured_top_k: int | None,
+) -> None:
+  memory_service = VertexAiRagMemoryService(
+      rag_corpus="unused",
+      similarity_top_k=configured_top_k,
+  )
+  fake_client = mocker.Mock()
+  fake_client.rag.retrieve_contexts.return_value = SimpleNamespace(
+      contexts=SimpleNamespace(contexts=[])
+  )
+  mocker.patch("agentplatform.Client", return_value=fake_client)
+
+  await memory_service.search_memory(
+      app_name="demo", user_id="alice", query="memory"
+  )
+
+  fake_client.rag.retrieve_contexts.assert_called_once()
+  kwargs = fake_client.rag.retrieve_contexts.call_args.kwargs
+  assert kwargs["query"].similarity_top_k == configured_top_k
+  # retrieveContexts reads top-k from the query; sending it on the store
+  # would put an undefined field on the request.
+  assert kwargs["vertex_rag_store"].similarity_top_k is None
+
+
 @pytest.mark.asyncio
 async def test_search_memory_rejects_ambiguous_legacy_display_names(mocker):
   """Ensures dotted user IDs cannot match another user's legacy memory."""