Update app.py

app.py

@@ -17,8 +17,8 @@ from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain.embeddings import HuggingFaceEmbeddings
 from langchain.vectorstores import FAISS
 
-PDF_PATH = "C:/Users/VASU/Downloads/Sample HI Policy.pdf"
-CSV_PATH = "C:/Users/VASU/Downloads/RAG_Test_Questions.csv"
+PDF_PATH = "Sample HI Policy.pdf"
+CSV_PATH = "RAG_Test_Questions.csv"
 
 st.set_page_config(page_title="PolicyGaido - Insurance Q&A", page_icon="📝", layout="wide")
 st.title("Insurance Policy Q&A Assistant")
@@ -32,7 +32,7 @@ if "initialized" not in st.session_state:
 
 with st.sidebar:
     st.header("Configuration")
-    model_option = st.selectbox("Select Language Model", ["BERT-for-QA", "DistilBERT-for-QA"])
+    model_option = st.selectbox("Select Language Model", ["BERT-for-QA"])
     device = "cuda" if torch.cuda.is_available() else "cpu"
     st.caption(f"Running on: {device}")
     
@@ -110,10 +110,9 @@ def get_answer(question, model_name):
     qa_pipeline = pipeline("question-answering", model="deepset/bert-base-cased-squad2" if model_name == "BERT-for-QA" else "distilbert-base-cased-distilled-squad", tokenizer="deepset/bert-base-cased-squad2", device=0 if torch.cuda.is_available() else -1)
     result = qa_pipeline(question=question, context=context)
     
-    # Calculate semantic similarity between question and context as a relevance proxy
     embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2", model_kwargs={'device': device})
     question_embedding = embeddings.embed_query(question)
-    context_embedding = embeddings.embed_query(context[:1000])  # Use first 1000 chars to avoid token limits
+    context_embedding = embeddings.embed_query(context[:1000])  # We are using the first 1000 chars only to avoid token limits
     
     # Compute cosine similarity
     similarity = np.dot(question_embedding, context_embedding) / (np.linalg.norm(question_embedding) * np.linalg.norm(context_embedding))
@@ -122,10 +121,8 @@ def get_answer(question, model_name):
     return result["answer"], docs, result["score"], relevance_score
 
 def evaluate_answer(answer, docs, confidence, relevance):
-    # Count potentially hallucinatory indicators
     hallucination_indicators = 0
     
-    # Check if answer contains content not found in supporting docs
     answer_found = False
     answer_words = set(answer.lower().split())
     
@@ -133,24 +130,22 @@ def evaluate_answer(answer, docs, confidence, relevance):
         for doc in docs:
             doc_content = doc.page_content.lower()
             overlap_count = sum(1 for word in answer_words if word in doc_content)
-            if overlap_count / len(answer_words) > 0.3:  # At least 30% of answer words are in document
+            if overlap_count / len(answer_words) > 0.3:  
                 answer_found = True
                 break
     
-    if not answer_found and len(answer_words) > 3:  # Only count if the answer is substantive
+    if not answer_found and len(answer_words) > 3:  
         hallucination_indicators += 1
-    
-    # Check for hedging language that might indicate uncertainty
+
     hedging_phrases = ["i think", "probably", "likely", "may", "might", "could be", "possibly", "perhaps"]
     if any(phrase in answer.lower() for phrase in hedging_phrases):
         hallucination_indicators += 1
     
-    # Return hallucination risk score (0-100)
     hallucination_risk = min(100, hallucination_indicators * 50)
     
     return {
         "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
-        "confidence": confidence * 100,  # Convert to percentage
+        "confidence": confidence * 100, 
         "relevance": relevance,
         "hallucination_risk": hallucination_risk
     }
@@ -179,10 +174,10 @@ with col2:
     st.subheader("Answer")
     if "last_answer" in st.session_state:
         question, answer, docs, evaluation = st.session_state["last_answer"]
-        st.markdown(f"**Question:** {question}")
-        st.markdown(f"**Answer:** {answer}")
+        st.markdown(f"Question: {question}")
+        st.markdown(f"Answer: {answer}")
+        
         
-        # Display evaluation metrics
         col_a, col_b, col_c = st.columns(3)
         with col_a:
             st.metric("Confidence", f"{evaluation['confidence']:.1f}%", 
@@ -197,16 +192,16 @@ with col2:
         
         with st.expander("View Source Information"):
             for i, doc in enumerate(docs):
-                st.markdown(f"**Source {i+1}:** {doc.page_content[:500]}...")
+                st.markdown(f"Source {i+1}: {doc.page_content[:500]}...")
+
 
-# History and statistics section
 st.divider()
 st.subheader("Evaluation History")
 
 if st.session_state.evaluation_history:
     history_df = pd.DataFrame(st.session_state.evaluation_history)
     
-    # Display summary statistics
+    # Displaying the  summary statistics
     st.subheader("Performance Statistics")
     col1, col2, col3 = st.columns(3)
     with col1:
@@ -216,7 +211,7 @@ if st.session_state.evaluation_history:
     with col3:
         st.metric("Avg. Hallucination Risk", f"{history_df['hallucination_risk'].mean():.1f}%")
     
-    # Show history table
+    # Here we are showing the history table
     st.dataframe(history_df)
 else:
     st.info("No evaluation history available yet. Ask some questions to build history.")