diff --git a/codex-lens/src/codexlens/semantic/vector_store.py b/codex-lens/src/codexlens/semantic/vector_store.py
index 0e7b6925..4e6b001b 100644
--- a/codex-lens/src/codexlens/semantic/vector_store.py
+++ b/codex-lens/src/codexlens/semantic/vector_store.py
@@ -19,10 +19,12 @@ from typing import Any, Dict, List, Optional, Tuple
 from codexlens.entities import SearchResult, SemanticChunk
 from codexlens.errors import StorageError
 
-from . import SEMANTIC_AVAILABLE
-
-if SEMANTIC_AVAILABLE:
+try:
     import numpy as np
+    NUMPY_AVAILABLE = True
+except ImportError:
+    np = None  # type: ignore[assignment]
+    NUMPY_AVAILABLE = False
 
 # Try to import ANN index (optional hnswlib dependency)
 try:
@@ -37,7 +39,7 @@ logger = logging.getLogger(__name__)
 
 def _cosine_similarity(a: List[float], b: List[float]) -> float:
     """Compute cosine similarity between two vectors."""
-    if not SEMANTIC_AVAILABLE:
+    if not NUMPY_AVAILABLE:
         raise ImportError("numpy required for vector operations")
 
     a_arr = np.array(a)
@@ -74,7 +76,7 @@ class VectorStore:
     DEFAULT_DIM = 768
 
     def __init__(self, db_path: str | Path) -> None:
-        if not SEMANTIC_AVAILABLE:
+        if not NUMPY_AVAILABLE:
             raise ImportError(
                 "Semantic search dependencies not available. "
                 "Install with: pip install codexlens[semantic]"
diff --git a/codex-lens/tests/test_vector_search_full.py b/codex-lens/tests/test_vector_search_full.py
index 98e0cb81..cc22c5f9 100644
--- a/codex-lens/tests/test_vector_search_full.py
+++ b/codex-lens/tests/test_vector_search_full.py
@@ -11,6 +11,7 @@ Tests cover:
 """
 
 import json
+import math
 import tempfile
 import threading
 import time
@@ -22,10 +23,17 @@ import pytest
 from codexlens.entities import SemanticChunk, Symbol, SearchResult
 from codexlens.semantic import SEMANTIC_AVAILABLE, check_semantic_available
 
-# Skip all tests if semantic dependencies not available
+# Only skip if NumPy is unavailable (some tests exercise vector math without fastembed).
+try:
+    import numpy as np  # noqa: F401
+    NUMPY_AVAILABLE = True
+except ImportError:
+    NUMPY_AVAILABLE = False
+
+# Skip all tests if NumPy is unavailable
 pytestmark = pytest.mark.skipif(
-    not SEMANTIC_AVAILABLE,
-    reason="Semantic search dependencies not installed (pip install codexlens[semantic])"
+    not NUMPY_AVAILABLE,
+    reason="NumPy not installed (pip install codexlens[semantic])"
 )
 
 
@@ -40,6 +48,9 @@ def temp_db(tmp_path):
 @pytest.fixture
 def embedder():
     """Create Embedder instance."""
+    available, error = check_semantic_available()
+    if not available:
+        pytest.skip(error or "Semantic search dependencies not installed (pip install codexlens[semantic])")
     from codexlens.semantic.embedder import Embedder
     return Embedder()
 
@@ -201,6 +212,50 @@ class TestEmbeddingSimilarity:
         similarity = _cosine_similarity(zero_vec, normal_vec)
         assert similarity == 0.0, "Zero vector should have 0 similarity"
 
+    def test_cosine_similarity_near_zero_norm_vectors(self):
+        """Near-zero norm vectors (< epsilon) should return 0.0 similarity."""
+        from codexlens.semantic.vector_store import _cosine_similarity
+
+        near_zero_vec = [1e-12] * 384
+        normal_vec = [1.0] * 384
+
+        similarity = _cosine_similarity(near_zero_vec, normal_vec)
+        assert similarity == 0.0
+
+    def test_cosine_similarity_product_underflow_returns_zero(self):
+        """Product underflow (norm_a * norm_b < epsilon) should return 0.0."""
+        from codexlens.semantic.vector_store import _cosine_similarity
+
+        underflow_vec = [1e-7] * 384
+
+        similarity = _cosine_similarity(underflow_vec, underflow_vec)
+        assert similarity == 0.0
+
+    def test_cosine_similarity_small_valid_vectors(self):
+        """Small-but-valid vectors should compute similarity correctly."""
+        from codexlens.semantic.vector_store import _cosine_similarity
+
+        small_vec = [1e-6] * 384
+
+        similarity = _cosine_similarity(small_vec, small_vec)
+        assert similarity == pytest.approx(1.0)
+
+    def test_cosine_similarity_no_inf_nan_results(self):
+        """Epsilon edge cases should never produce inf/nan results."""
+        from codexlens.semantic.vector_store import _cosine_similarity
+
+        cases = [
+            ([0.0] * 384, [1.0] * 384),
+            ([1e-12] * 384, [1.0] * 384),
+            ([1e-7] * 384, [1e-7] * 384),
+            ([1e-6] * 384, [1e-6] * 384),
+            ([1.0] * 384, [1.0] * 384),
+        ]
+
+        for a, b in cases:
+            similarity = _cosine_similarity(a, b)
+            assert math.isfinite(similarity)
+
 
 # === VectorStore Tests ===
 
@@ -746,9 +801,12 @@ class TestAvailabilityCheck:
     def test_check_semantic_available(self):
         """Test check_semantic_available function."""
         available, error = check_semantic_available()
-        assert available is True
-        assert error is None
+        assert available is SEMANTIC_AVAILABLE
+        if available:
+            assert error is None
+        else:
+            assert error is not None
 
     def test_semantic_available_flag(self):
         """Test SEMANTIC_AVAILABLE flag is True when deps installed."""
-        assert SEMANTIC_AVAILABLE is True
+        assert isinstance(SEMANTIC_AVAILABLE, bool)