From 8d45fb3c9aed8152a07d14e2ee1e5ebf0be05f96 Mon Sep 17 00:00:00 2001
From: Aari <ansusu1025@gmail.com>
Date: Fri, 31 Jul 2026 23:02:27 +0800
Subject: [PATCH] fix(elasticsearch): set size on KNN search to respect top_k
 (#5910)

---
 mem0/vector_stores/elasticsearch.py       |  5 ++-
 tests/vector_stores/test_elasticsearch.py | 39 +++++++++++++++++++++++
 2 files changed, 43 insertions(+), 1 deletion(-)

diff --git a/mem0/vector_stores/elasticsearch.py b/mem0/vector_stores/elasticsearch.py
index 96e520236..6578e9863 100644
--- a/mem0/vector_stores/elasticsearch.py
+++ b/mem0/vector_stores/elasticsearch.py
@@ -163,7 +163,10 @@ class ElasticsearchDB(VectorStoreBase):
             search_query = self.custom_search_query(vectors, top_k, filters)
         else:
             search_query = {
-                "knn": {"field": "vector", "query_vector": vectors, "k": top_k, "num_candidates": top_k * 2}
+                # Without `size`, Elasticsearch caps the response at its default of 10 hits
+                # regardless of `knn.k`, silently truncating results when top_k > 10.
+                "size": top_k,
+                "knn": {"field": "vector", "query_vector": vectors, "k": top_k, "num_candidates": top_k * 2},
             }
             if filters:
                 filter_conditions = []
diff --git a/tests/vector_stores/test_elasticsearch.py b/tests/vector_stores/test_elasticsearch.py
index a537ccd6a..57bc9d81f 100644
--- a/tests/vector_stores/test_elasticsearch.py
+++ b/tests/vector_stores/test_elasticsearch.py
@@ -209,6 +209,8 @@ class TestElasticsearchDB(unittest.TestCase):
         self.assertEqual(body["knn"]["query_vector"], vectors)
         self.assertEqual(body["knn"]["k"], 5)
         self.assertEqual(body["knn"]["num_candidates"], 10)
+        # `size` must be set or ES caps the response at its default of 10 hits.
+        self.assertEqual(body["size"], 5)
 
         # Verify results
         self.assertEqual(len(results), 1)
@@ -216,6 +218,43 @@ class TestElasticsearchDB(unittest.TestCase):
         self.assertEqual(results[0].score, 0.8)
         self.assertEqual(results[0].payload, {"key1": "value1"})
 
+    def test_search_sets_size_to_top_k(self):
+        """Regression for #5909: KNN search must set `size=top_k` or Elasticsearch caps
+        the response at its default of 10 hits regardless of `knn.k`.
+
+        Memory._search_vector_store over-fetches a scoring pool of `max(limit * 4, 60)`
+        candidates for hybrid re-ranking, so even a default search(top_k=20) reaches this
+        layer with top_k=80. Without `size`, that pool is silently truncated to 10.
+        """
+        self.client_mock.search.return_value = {"hits": {"hits": []}}
+
+        # internal_limit in Memory._search_vector_store for a default search(top_k=20)
+        over_fetch = max(20 * 4, 60)
+        self.es_db.search(query="", vectors=[[0.1] * 1536], top_k=over_fetch)
+
+        body = self.client_mock.search.call_args[1]["body"]
+        self.assertEqual(body["size"], over_fetch)
+        self.assertEqual(body["knn"]["k"], over_fetch)
+        self.assertGreater(body["size"], 10)
+
+    def test_search_with_filters_keeps_size(self):
+        """Filters must not displace the response size set for KNN search."""
+        self.client_mock.search.return_value = {"hits": {"hits": []}}
+
+        self.es_db.search(
+            query="",
+            vectors=[[0.1] * 1536],
+            top_k=25,
+            filters={"user_id": "u1"},
+        )
+
+        body = self.client_mock.search.call_args[1]["body"]
+        self.assertEqual(body["size"], 25)
+        self.assertEqual(
+            body["knn"]["filter"],
+            {"bool": {"must": [{"term": {"metadata.user_id": "u1"}}]}},
+        )
+
     def test_custom_search_query(self):
         # Mock custom search query
         self.es_db.custom_search_query = Mock()