diff --git a/rabitqlib/quantization/rabitq_impl.hpp b/rabitqlib/quantization/rabitq_impl.hpp
index af0371d..ff2ac8a 100644
--- a/rabitqlib/quantization/rabitq_impl.hpp
+++ b/rabitqlib/quantization/rabitq_impl.hpp
@@ -100,6 +100,11 @@ inline void one_bit_code_with_factor(
     // dot product between centroid and xu_cb
     T ip_cent_xucb = dot_product<T>(centroid, xu_cb.data(), dim);
 
+    // corner case
+    if (ip_resi_xucb == 0) {
+        ip_resi_xucb = std::numeric_limits<T>::infinity();
+    }
+
     // We use unnormalized vector to get error factor. To be more specific,
     // sqrt((1 - <o, o_bar>^2) / <o, o_bar>^2) / sqrt(dim - 1) = 3rd item in following
     // expression
@@ -465,6 +470,11 @@ inline void ex_bits_code_with_factor(
     T ip_resi_xucb = dot_product<T>(residual_arr.data(), xu_cb.data(), dim);
     T ip_cent_xucb = dot_product<T>(centroid, xu_cb.data(), dim);
 
+    // corner case
+    if (ip_resi_xucb == 0) {
+        ip_resi_xucb = std::numeric_limits<T>::infinity();
+    }
+
     T tmp_error =
         l2_norm * kConstEpsilon *
         std::sqrt(
@@ -556,7 +566,8 @@ static inline void rabitq_scalar_impl(
 
     float norm_data = std::sqrt(l2norm_sqr(residual_arr.data(), dim));
     float norm_quan = std::sqrt(l2norm_sqr(u_cb.data(), dim));
-    float cos_similarity = dot_product<T>(residual_arr.data(), u_cb.data(), dim) / (norm_data * norm_quan);
+    float cos_similarity =
+        dot_product<T>(residual_arr.data(), u_cb.data(), dim) / (norm_data * norm_quan);
 
     if (scalar_quantizer_type == ScalarQuantizerType::RECONSTRUCTION) {
         delta = norm_data / norm_quan * cos_similarity;