fix(storage): normalize Milvus cosine scores

Signed-off-by: Cheney Zhang <chen.zhang@zilliz.com>

Author: zc277584121

openviking/storage/vectordb_adapters/milvus_adapter.py

@@ -205,6 +205,8 @@ def _score_from_hit(hit: Dict[str, Any], distance_metric: str) -> float:
         return 0.0
     if not math.isfinite(score):
         return 0.0
+    if distance_metric == "cosine":
+        return 1.0 - score
     if distance_metric == "l2":
         return 1.0 / (1.0 + max(score, 0.0))
     return score

tests/storage/test_milvus_adapter.py

@@ -198,6 +198,13 @@ def test_scope_roots_encoding_is_token_safe():
     assert "\n/a/c\n" not in encoded
 
 
+def test_score_from_cosine_distance_is_higher_is_better():
+    from openviking.storage.vectordb_adapters.milvus_adapter import _score_from_hit
+
+    assert _score_from_hit({"distance": 0.0}, "cosine") == pytest.approx(1.0)
+    assert _score_from_hit({"distance": 1.0}, "cosine") == pytest.approx(0.0)
+
+
 class _FakeSchema:
     def __init__(self) -> None:
         self.fields = []
@@ -358,6 +365,7 @@ def _new_adapter() -> MilvusCollectionAdapter:
             output_fields=["id", "uri", "abstract", "level"],
         )
         assert [item["id"] for item in result] == ["doc-1"]
+        assert result[0]["_score"] == pytest.approx(1.0)
         assert adapter.count(Eq("account_id", "missing")) == 0
         assert adapter.count() == 2
         assert adapter.delete(ids=["doc-2"]) == 1