fix(reranker): score HuggingFace reranker with sigmoid, not min-max (#5715)

mem0/reranker/huggingface_reranker.py

@@ -59,6 +59,26 @@ class HuggingFaceReranker(BaseReranker):
         self.model.to(self.device)
         self.model.eval()
 
+    @staticmethod
+    def _normalize_scores(scores: List[float]) -> List[float]:
+        """Map raw cross-encoder logits into the [0, 1] range via a sigmoid.
+
+        Cross-encoder rerankers (e.g. ``BAAI/bge-reranker-*``) emit unbounded
+        logits; the documented way to obtain an interpretable [0, 1] relevance
+        score is a per-document sigmoid (``1 / (1 + e^-x)``), which preserves the
+        ranking order.
+
+        This replaces the previous min-max scaling, which produced *set-relative*
+        scores: the lowest-ranked document was always forced to 0.0, and a single
+        document (or any set of tied scores) collapsed to 0.0 — wrongly reporting
+        a result as completely irrelevant. Sigmoid scores each document on its own
+        merit, so those cases are handled naturally.
+        """
+        if not scores:
+            return []
+        arr = np.asarray(scores, dtype=float)
+        return (1.0 / (1.0 + np.exp(-arr))).tolist()
+
     def rerank(self, query: str, documents: List[Dict[str, Any]], top_k: int = None) -> List[Dict[str, Any]]:
         """
         Rerank documents using HuggingFace cross-encoder model.
@@ -118,9 +138,7 @@ class HuggingFaceReranker(BaseReranker):
 
             # Normalize scores if requested
             if self.config.normalize:
-                scores = np.array(scores)
-                scores = (scores - scores.min()) / (scores.max() - scores.min() + 1e-8)
-                scores = scores.tolist()
+                scores = self._normalize_scores(scores)
 
             # Combine documents with scores
             doc_score_pairs = list(zip(documents, scores))