Create rrf_safe_similarity.py

rrf_safe_similarity.py

@@ -0,0 +1,48 @@
+import numpy as np
+
+
+def stabilize_vector(
+    vec: np.ndarray,
+    r_min: float = 1e-4,
+    r_max: float = 1e4,
+):
+    """
+    Applies RRF logarithmic norm stabilization.
+    Returns (stabilized_vector, audit_event | None)
+    """
+
+    norm = np.linalg.norm(vec)
+
+    audit_event = None
+
+    if norm < r_min:
+        scale = r_min / (norm + 1e-12)
+        vec = vec * scale
+
+        audit_event = {
+            "type": "stabilization_triggered",
+            "reason": "norm_below_min",
+            "original_norm": float(norm),
+            "target_norm": r_min,
+        }
+
+    elif norm > r_max:
+        scale = r_max / norm
+        vec = vec * scale
+
+        audit_event = {
+            "type": "stabilization_triggered",
+            "reason": "norm_above_max",
+            "original_norm": float(norm),
+            "target_norm": r_max,
+        }
+
+    return vec, audit_event
+
+
+def safe_cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
+    """
+    Cosine similarity with numerical safety.
+    """
+    denom = (np.linalg.norm(a) * np.linalg.norm(b)) + 1e-12
+    return float(np.dot(a, b) / denom)