Upload doc_searcher_v2.py with huggingface_hub

doc_searcher_v2.py

@@ -14,17 +14,30 @@ class DocSearcherV2:
         self.sparse_model = SparseTextEmbedding(SPARSE_MODEL)
         self.qdrant_client = QdrantClient(QDRANT_URL,api_key=QDRANT_API_KEY,timeout=30)
     
-    async def search_semantic(self, text: str):
+    async def search_semantic(self, text: str, qdrant_limit: int = 20, top_k: int = 3):
+        """
+        Semantička pretraga sa reranking-om
         
+        Args:
+            text: Query tekst
+            qdrant_limit: Broj chunk-ova iz Qdrant-a (default: 20)
+            top_k: Broj najboljih rezultata nakon rerankinga (default: 3)
+        """
         queries = [text]
         dense_query = self.model.encode(text).tolist()
         sparse_query = next(self.sparse_model.query_embed(text))
 
+        # Hybrid search: dense + sparse za bolje rezultate
         prefetch = [
                 models.Prefetch(
                     query=dense_query,
                     using="Qwen/Qwen3-Embedding-0.6B",
-                    limit=100
+                    limit=qdrant_limit
+                ),
+                models.Prefetch(
+                    query=models.SparseVector(**sparse_query.as_object()),
+                    using=SPARSE_MODEL,
+                    limit=qdrant_limit
                 ),
         ]
             
@@ -35,7 +48,7 @@ class DocSearcherV2:
                 fusion=models.Fusion.RRF,
             ),
             with_payload=True,
-            limit = 100,
+            limit=qdrant_limit,
         ).points
 
         # Sačuvaj kompletan hit objekat i tekst za reranking
@@ -52,9 +65,12 @@ class DocSearcherV2:
         if not texts_for_reranking:
             return []
             
-        # Reranking - vraća top_10 sa (score, query, document) tuple-ovima
-        # queries * len(...) znači da svaki dokument dobija isti query
-        reranked_results = self.reranker.compute_logits(queries * len(texts_for_reranking), texts_for_reranking)
+        # Reranking - ISPRAVLJENO: jedan query za sve dokumente
+        reranked_results = self.reranker.compute_logits(
+            queries * len(texts_for_reranking),  # Svaki dokument dobija isti query
+            texts_for_reranking,
+            top_k=top_k  # Vrati samo top_k rezultata
+        )
         
         # Kombinuj rezultate: mapiraj rerank skorove sa originalnim hit-ovima
         # reranked_results je lista tuple-ova: [(score, query, document_text), ...]