.

RagWithConfidenceScore.py

@@ -16,7 +16,7 @@ from concurrent.futures import ThreadPoolExecutor
 class RagWithScore:
     def __init__(self, model_name="sentence-transformers/all-MiniLM-L6-v2", 
                 cross_encoder_name="cross-encoder/ms-marco-TinyBERT-L-2-v2",
-                llm_name="gpt2",
+                llm_name="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
                 documents_dir="financial_docs"):
         """
         Initialize the Financial RAG system
@@ -39,12 +39,12 @@ class RagWithScore:
         "text-generation",
         model=llm_name,
         tokenizer=self.tokenizer,
-        torch_dtype=torch.bfloat16,  # Use float16 if bfloat16 is not supported
+        torch_dtype=torch.bfloat16,
         device_map="auto",
-        max_new_tokens=512,          # Adjust based on your needs
-        do_sample=False,             # Set to False for deterministic outputs
-        temperature=0.2,             # Reduce randomness
-        top_p=1.0                    # No nucleus sampling
+        max_new_tokens=512,
+        do_sample=False,  # Set to False for deterministic outputs
+        temperature=0.2,   # Reduce randomness
+        top_p=1.0          # No nucleus sampling
 )
         
         # Store paths
@@ -71,6 +71,7 @@ class RagWithScore:
 
     import os
 
+## Loadung document and creating vector index at the start of the  application
     def load_and_process_documents(self):
         """Load, split and process financial documents"""
 
@@ -91,19 +92,8 @@ class RagWithScore:
     
         return self.vector_store
 
-    def load_or_create_vector_store(self):
-        try:
-            print("Loading existing FAISS index...")
-            self.vector_store = FAISS.load_local("faiss_index", self.embedding_model)
-            print("FAISS index loaded successfully")
-        except Exception as e:
-            print(f"Error loading FAISS index: {e}")
-            print("Creating new FAISS index...")
-            # Code to create a new vector store
-            documents = self.load_and_process_documents()  # Make sure this method exists
-            print("New FAISS index created and saved")
-
 
+## generating response  with the query and context by the help of the prompt and calling the slm with the prompt
     def generate_answer(self, query, context):
         """Generate answer and calculate confidence score concurrently."""
         # Format context into a single string
@@ -142,22 +132,8 @@ class RagWithScore:
         return answer
 
 
-    # def calculate_confidence_score(self, query, retrieved_docs, answer):
-    #     """A simpler confidence score calculation focused on consistency and LLM confidence"""
-        
-    #     # Get LLM confidence
-    #     llm_confidence = self._get_llm_confidence(query, retrieved_docs, answer)
-        
-    #     # Get consistency score
-    #     consistency_score = self._measure_answer_consistency(query, retrieved_docs, answer)
-        
-    #     # Simple weighted average
-    #     confidence_score = (0.6 * consistency_score) + (0.4 * llm_confidence)
-
-    #     print(f"confidence score : {confidence_score}")
-        
-    #     return confidence_score
 
+## for  confidence score cosine  similarity is  calculated  between the query  embedding and answer embedding
     def calculate_confidence_score(self, query, retrieved_docs, answer):
         """
         Calculate confidence score using embedding similarity (parallelized).
@@ -175,6 +151,7 @@ class RagWithScore:
         return similarity
     
    
+   ## confidence level is determined from the confidence score
     def get_confidence_level(self, confidence_score):
         """
         Convert numerical confidence score to a level (high, medium, low)
@@ -194,6 +171,7 @@ class RagWithScore:
         else:
             return "very low"
 
+## guardrail is  applied to filter harmful user  queries
     def apply_input_guardrail(self, query):
         """Check if query violates input guardrails"""
         query_lower = query.lower()
@@ -204,6 +182,7 @@ class RagWithScore:
         
         return False, ""
 
+## first the   to 5 chunks are retrieved. then after reranking with cross encoder top 2 are rerieved
     def retrieve_with_reranking(self, query, top_k=5, rerank_top_k=3):
 
         print("retrieve_with_reranking start")
@@ -241,6 +220,7 @@ class RagWithScore:
 
         return [doc for (doc, _), _ in reranked_results[:rerank_top_k]]
 
+## to handle irrerelevant questions, a rule based claasifier is  bein used to classify the questions
     def is_financial_question(self,query):
         financial_keywords = [
             "finance", "financial", "revenue", "profit", "loss", "ebitda", "cash flow",
@@ -251,7 +231,8 @@ class RagWithScore:
         ]
         query_lower = query.lower()
         return any(keyword in query_lower for keyword in financial_keywords)
-            
+
+    ##the   pipeline of answer and confidence score generation from the query        
     def answer_question(self, query):
         """End-to-end pipeline to answer a question with confidence score"""