diff --git a/include/svs/core/distance/cosine.h b/include/svs/core/distance/cosine.h
index c8e41e7..53c9ec6 100644
--- a/include/svs/core/distance/cosine.h
+++ b/include/svs/core/distance/cosine.h
@@ -306,9 +306,26 @@ template <size_t N> struct CosineSimilarityImpl<N, float, int8_t> {
 template <size_t N> struct CosineSimilarityImpl<N, float, Float16> {
     SVS_NOINLINE static float
     compute(const float* a, const Float16* b, float a_norm, lib::MaybeStatic<N> length) {
-        auto [sum, norm] = simd::generic_simd_op(CosineFloatOp<16>(), a, b, length);
+        auto [sum, norm] = simd::generic_simd_op(CosineFloatOp<16>{}, a, b, length);
+        return sum / (std::sqrt(norm) * a_norm);
+    }
+};
+
+template <size_t N> struct CosineSimilarityImpl<N, Float16, float> {
+    SVS_NOINLINE static float
+    compute(const Float16* a, const float* b, float a_norm, lib::MaybeStatic<N> length) {
+        auto [sum, norm] = simd::generic_simd_op(CosineFloatOp<16>{}, a, b, length);
         return sum / (std::sqrt(norm) * a_norm);
     }
 };
+
+template <size_t N> struct CosineSimilarityImpl<N, Float16, Float16> {
+    SVS_NOINLINE static float
+    compute(const Float16* a, const Float16* b, float a_norm, lib::MaybeStatic<N> length) {
+        auto [sum, norm] = simd::generic_simd_op(CosineFloatOp<16>{}, a, b, length);
+        return sum / (std::sqrt(norm) * a_norm);
+    }
+};
+
 #endif
 } // namespace svs::distance
diff --git a/include/svs/index/inverted/common.h b/include/svs/index/inverted/common.h
index 88c6625..6bc3395 100644
--- a/include/svs/index/inverted/common.h
+++ b/include/svs/index/inverted/common.h
@@ -40,7 +40,12 @@ template <typename T> inline T bound_with(T nearest, T epsilon, svs::DistanceL2)
 }
 
 template <typename T> inline T bound_with(T nearest, T epsilon, svs::DistanceIP) {
-    // TODO: What do we do if the best match is simply bad?
+    assert(nearest > 0.0f);
+    return nearest / (1 + epsilon);
+}
+
+template <typename T>
+inline T bound_with(T nearest, T epsilon, svs::DistanceCosineSimilarity) {
     assert(nearest > 0.0f);
     return nearest / (1 + epsilon);
 }