Upload 2 files

config.py

@@ -19,4 +19,5 @@ CHROMA_COLLECTION_NAME: str = "faq"
 SIMILARITY_THRESHOLD: float = float(os.getenv("SIMILARITY_THRESHOLD", "0.80"))
 
 TOP_K = 3
-CROSS_ENCODER_MODEL_NAME = "cross-encoder/mmarco-mMiniLMv2-L12-H384-v1"
+CROSS_ENCODER_MODEL_NAME = "cross-encoder/mmarco-mMiniLMv2-L12-H384-v1"
+CROSS_ENCODER_THRESHOLD: float = float(os.getenv("CROSS_ENCODER_THRESHOLD", "0.20"))

embeddings.py

@@ -23,11 +23,8 @@ import chromadb
 from sentence_transformers import SentenceTransformer, CrossEncoder
 import torch
 from config import (
-    CHROMA_COLLECTION_NAME,
-    MODEL_NAME,
-    SIMILARITY_THRESHOLD,
-    TOP_K,
-    CROSS_ENCODER_MODEL_NAME
+    CHROMA_COLLECTION_NAME, MODEL_NAME, SIMILARITY_THRESHOLD,
+    TOP_K, CROSS_ENCODER_MODEL_NAME, CROSS_ENCODER_THRESHOLD
 )
 from faq_loader import FAQEntry
 
@@ -204,6 +201,7 @@ class EmbeddingManager:
         # Récupérer la question et les paramètres depuis le payload
         question = payload.question
         similarity_threshold = payload.similarity_threshold if payload.similarity_threshold is not None else SIMILARITY_THRESHOLD
+        cross_encoder_threshold = payload.cross_encoder_threshold if payload.cross_encoder_threshold is not None else CROSS_ENCODER_THRESHOLD
         top_k = getattr(payload, "top_k", None) or TOP_K
 
         # Générer l'embedding de la question
@@ -265,12 +263,22 @@ class EmbeddingManager:
         for candidate, cross_score in zip(candidates, cross_scores):
             candidate.cross_encoder_score = float(cross_score)
 
-        # Sélectionner le meilleur score cross-encoder
-        best_candidate = max(
+        # Sélectionner le meilleur score cross-encoder et le second meilleur pour calculer l'écart entre les 2
+        ranked_candidates = sorted(
             candidates,
-            key=lambda c: c.cross_encoder_score if c.cross_encoder_score is not None else float("-inf"),
+            key=lambda c: c.cross_encoder_score
+            if c.cross_encoder_score is not None
+            else float("-inf"),
+            reverse=True,
         )
 
+        best_candidate = ranked_candidates[0]
+        second_candidate = ranked_candidates[1] if len(ranked_candidates) > 1 else None
+
+        best_score = best_candidate.cross_encoder_score or 0.0
+        relative_gap = abs((best_score - second_candidate.cross_encoder_score) / best_score) \
+                       if second_candidate and best_score != 0 else 0
+
         logger.info(
             "Candidats re-classés par cross-encoder: "
             + ", ".join(
@@ -278,10 +286,9 @@ class EmbeddingManager:
             )
         )
 
-        """
-        if best_candidate.cross_encoder_score is None or best_candidate.cross_encoder_score < cross_encoder_threshold:
+        if best_candidate.cross_encoder_score is None or relative_gap < cross_encoder_threshold:
             logger.info(
-                f"Meilleur score cross-encoder {best_candidate.cross_encoder_score:.3f} "
+                f"GAP entre 1er et 2ème meilleurs scores: {relative_gap:.3f} "
                 f"< seuil {cross_encoder_threshold}. Question: {question}"
             )
             return AnswerOutput(
@@ -294,7 +301,7 @@ class EmbeddingManager:
                 list_candidates=candidates,
                 confidence=False,
             )
-        """
+
 
         logger.info(
             f"Match trouvé - Theme: {best_candidate.theme}, "