From 6486c5685035264b388e4c8b38091e8062f3cc0f Mon Sep 17 00:00:00 2001
From: catlog22 <catlog22@github.com>
Date: Sun, 28 Dec 2025 21:11:30 +0800
Subject: [PATCH] fix(vector-store): add epsilon tolerance for norm checks

Replace exact zero comparison with epsilon-based check (< 1e-10)
in _cosine_similarity to handle floating point precision issues.
Also check for product underflow to prevent inf/nan from division
by very small numbers.

Solution-ID: SOL-20251228113619
Issue-ID: ISS-1766921318981-11
Task-ID: T1
---
 codex-lens/src/codexlens/semantic/vector_store.py | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/codex-lens/src/codexlens/semantic/vector_store.py b/codex-lens/src/codexlens/semantic/vector_store.py
index c7133e42..0e7b6925 100644
--- a/codex-lens/src/codexlens/semantic/vector_store.py
+++ b/codex-lens/src/codexlens/semantic/vector_store.py
@@ -46,10 +46,16 @@ def _cosine_similarity(a: List[float], b: List[float]) -> float:
     norm_a = np.linalg.norm(a_arr)
     norm_b = np.linalg.norm(b_arr)
 
-    if norm_a == 0 or norm_b == 0:
+    EPSILON = 1e-10
+    # Use epsilon tolerance to avoid division by (near-)zero due to floating point precision.
+    if norm_a < EPSILON or norm_b < EPSILON:
         return 0.0
 
-    return float(np.dot(a_arr, b_arr) / (norm_a * norm_b))
+    denom = norm_a * norm_b
+    if denom < EPSILON:
+        return 0.0
+
+    return float(np.dot(a_arr, b_arr) / denom)
 
 
 class VectorStore: