Added retrieval timings logs

rag.py

@@ -85,9 +85,16 @@ class RAG:
         """Retrieve the most relevant contexts for a given instruction."""
         
         logging.info("RETRIEVE DOCUMENTS")
-        documents_retrieved = self.vectore_store.similarity_search_with_score(instruction, k=self.rerank_number_contexts)
+        logging.info(f"Instruction: {instruction}")
+        embedding = self.vectore_store._embed_query(instruction)
+        logging.info(f"Query embedding generated: {len(embedding)}")
+        documents_retrieved = self.vectore_store.similarity_search_with_score_by_vector(
+            embedding, 
+            k=self.rerank_number_contexts)
         logging.info(f"Documents retrieved: {len(documents_retrieved)}")
 
+        # documents_retrieved = self.vectore_store.similarity_search_with_score(instruction, k=self.rerank_number_contexts)
+
         if self.rerank_model:
             logging.info("RERANK DOCUMENTS")
             documents_reranked = self.rerank_contexts(instruction, documents_retrieved, number_of_contexts=number_of_contexts)