ok

app/api/v1/semantic_router.py

@@ -34,6 +34,7 @@ async def route_query(
         models=result.get("models", []),
         error=result.get("error"),
         confidence=result.get("confidence"),
+        margin=result.get("margin"),
         threshold=result.get("threshold"),
         matched_utterance=result.get("matched_utterance"),
     )

app/services/semantic_router_service.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import logging
+from collections import defaultdict
 from typing import Any, Optional
 
 import numpy as np
@@ -57,17 +58,25 @@ class SemanticRouterService:
         best_idx = int(np.argmax(similarities))
         best_score = float(similarities[best_idx])
 
-        sorted_scores = sorted(similarities, reverse=True)
-        margin = (sorted_scores[0] - sorted_scores[1]) if len(sorted_scores) > 1 else 1.0
+        # Route-level scoring: take the max score per route
+        route_scores: dict[int, list[float]] = defaultdict(list)
+        for i, score in enumerate(similarities):
+            route_scores[utterance_to_route[i]].append(float(score))
 
-        if best_score < threshold or margin < 0.01:
+        route_best_scores = {rid: max(scores) for rid, scores in route_scores.items()}
+        sorted_routes = sorted(route_best_scores.items(), key=lambda x: x[1], reverse=True)
+        best_route_score = sorted_routes[0][1]
+        second_best_route_score = sorted_routes[1][1] if len(sorted_routes) > 1 else 1.0
+        route_margin = best_route_score - second_best_route_score
+
+        if best_score < threshold or route_margin < 0.001:
             return {
                 "success": True,
                 "name": None,
                 "models": [],
                 "error": None,
                 "confidence": best_score,
-                "margin": margin,
+                "margin": route_margin,
                 "threshold": threshold,
                 "matched_utterance": all_utterances[best_idx],
             }
@@ -81,7 +90,7 @@ class SemanticRouterService:
             "models": matched_route.get("models", []),
             "error": None,
             "confidence": best_score,
-            "margin": margin,
+            "margin": route_margin,
             "threshold": threshold,
             "matched_utterance": all_utterances[best_idx],
         }