diff --git a/faiss/IndexHNSW.cpp b/faiss/IndexHNSW.cpp
index c0bb81c052..e18e203208 100644
--- a/faiss/IndexHNSW.cpp
+++ b/faiss/IndexHNSW.cpp
@@ -35,26 +35,6 @@
 #include <faiss/utils/random.h>
 #include <faiss/utils/sorting.h>
 
-extern "C" {
-
-/* declare BLAS functions, see http://www.netlib.org/clapack/cblas/ */
-
-int sgemm_(
-        const char* transa,
-        const char* transb,
-        FINTEGER* m,
-        FINTEGER* n,
-        FINTEGER* k,
-        const float* alpha,
-        const float* a,
-        FINTEGER* lda,
-        const float* b,
-        FINTEGER* ldb,
-        float* beta,
-        float* c,
-        FINTEGER* ldc);
-}
-
 namespace faiss {
 
 using MinimaxHeap = HNSW::MinimaxHeap;
@@ -340,7 +320,7 @@ void IndexHNSW::range_search(
         RangeSearchResult* result,
         const SearchParameters* params) const {
     using RH = RangeSearchBlockResultHandler<HNSW::C>;
-    RH bres(result, radius);
+    RH bres(result, is_similarity_metric(metric_type) ? -radius : radius);
 
     hnsw_search(this, n, x, bres, params);
 
diff --git a/tests/test_graph_based.py b/tests/test_graph_based.py
index d5797186da..a5dcd4c03a 100644
--- a/tests/test_graph_based.py
+++ b/tests/test_graph_based.py
@@ -184,6 +184,34 @@ def test_abs_inner_product(self):
         # 4769 vs. 500*10
         self.assertGreater(inter, Iref.size * 0.9)
  
+ 
+class Issue3684(unittest.TestCase):
+
+    def test_issue3684(self):
+        np.random.seed(1234)  # For reproducibility
+        d = 256  # Example dimension
+        nb = 10  # Number of database vectors
+        nq = 2   # Number of query vectors
+        xb = np.random.random((nb, d)).astype('float32')
+        xq = np.random.random((nq, d)).astype('float32')
+
+        faiss.normalize_L2(xb)  # Normalize both query and database vectors
+        faiss.normalize_L2(xq)
+
+        hnsw_index_ip = faiss.IndexHNSWFlat(256, 16, faiss.METRIC_INNER_PRODUCT)
+        hnsw_index_ip.hnsw.efConstruction = 512
+        hnsw_index_ip.hnsw.efSearch = 512
+        hnsw_index_ip.add(xb)
+
+        # test knn 
+        D, I = hnsw_index_ip.search(xq, 10)
+        self.assertTrue(np.all(D[:, :-1] >= D[:, 1:]))
+
+        # test range search 
+        radius = 0.74  # Cosine similarity threshold
+        lims, D, I = hnsw_index_ip.range_search(xq, radius)
+        self.assertTrue(np.all(D >= radius))
+
 
 class TestNSG(unittest.TestCase):