only semantic search top k = 30, cut off = 0.7

index_retriever.py

@@ -12,28 +12,26 @@ def create_vector_index(documents):
 
 def create_query_engine(vector_index):
     try:
-        # --- Semantic-only retriever ---
         vector_retriever = VectorIndexRetriever(
-            index=vector_index,
-            similarity_top_k=30,      # recommended default
-            similarity_cutoff=0.78    # filter weak matches
+            index=vector_index, 
+            similarity_top_k=30,
+            similarity_cutoff=0.7
         )
-
+        
         custom_prompt_template = PromptTemplate(PROMPT_SIMPLE_POISK)
-
         response_synthesizer = get_response_synthesizer(
             response_mode=ResponseMode.TREE_SUMMARIZE,
             text_qa_template=custom_prompt_template
         )
-
+        
         query_engine = RetrieverQueryEngine(
             retriever=vector_retriever,
             response_synthesizer=response_synthesizer
         )
-
-        log_message("Semantic-only query engine успешно создан")
+        
+        log_message("Query engine успешно создан (только векторный поиск)")
         return query_engine
-
+        
     except Exception as e:
         log_message(f"Ошибка создания query engine: {str(e)}")
-        raise
+        raise

utils.py

@@ -6,7 +6,7 @@ from llama_index.embeddings.huggingface import HuggingFaceEmbedding
 from sentence_transformers import CrossEncoder
 from config import AVAILABLE_MODELS, DEFAULT_MODEL, GOOGLE_API_KEY
 import time
-from index_retriever import rerank_nodes
+# from index_retriever import rerank_nodes
 from my_logging import log_message
 from config import PROMPT_SIMPLE_POISK
 
@@ -260,31 +260,11 @@ def answer_question(question, query_engine, reranker, current_model, chunks_df=N
         
         llm = get_llm_model(current_model)
         
-        query_variations = expand_query(question, llm)
+        retrieved_nodes = query_engine.retriever.retrieve(question)
         
-        all_nodes = []
-        seen_node_ids = set()
+        log_message(f"Получено {len(retrieved_nodes)} узлов")
         
-        for query_var in query_variations:
-            retrieved = query_engine.retriever.retrieve(query_var)
-            for node in retrieved:
-                node_id = f"{node.node_id if hasattr(node, 'node_id') else hash(node.text)}"
-                if node_id not in seen_node_ids:
-                    all_nodes.append(node)
-                    seen_node_ids.add(node_id)
-        
-        log_message(f"Получено {len(all_nodes)} уникальных узлов из {len(query_variations)} запросов")
-        
-        reranked_nodes = rerank_nodes(
-            question,  
-            all_nodes, 
-            reranker, 
-            top_k=20,
-            min_score_threshold=0.5, 
-            diversity_penalty=0.3
-        )
-        
-        formatted_context = format_context_for_llm(reranked_nodes)
+        formatted_context = format_context_for_llm(retrieved_nodes)
         
         enhanced_question = f"""Контекст из базы данных:
 {formatted_context}
@@ -301,7 +281,7 @@ def answer_question(question, query_engine, reranker, current_model, chunks_df=N
         
         log_message(f"Обработка завершена за {processing_time:.2f}с")
         
-        sources_html = generate_sources_html(reranked_nodes, chunks_df)
+        sources_html = generate_sources_html(retrieved_nodes, chunks_df)
         
         answer_with_time = f"""<div style='background-color: #2d3748; color: white; padding: 20px; border-radius: 10px; margin-bottom: 10px;'>
         <h3 style='color: #63b3ed; margin-top: 0;'>Ответ (Модель: {current_model}):</h3>
@@ -312,7 +292,7 @@ def answer_question(question, query_engine, reranker, current_model, chunks_df=N
         </div>"""
         
         chunk_info = []
-        for node in reranked_nodes:
+        for node in retrieved_nodes :
             metadata = node.metadata if hasattr(node, 'metadata') else {}
             chunk_info.append({
                 'document_id': metadata.get('document_id', 'unknown'),