Update flask_app.py

flask_app.py

@@ -109,28 +109,56 @@ import re
 #         return "I don't know."
 
 # Function to perform model inference
+import re
+from fuzzywuzzy import fuzz  # For relevance matching
+
+def is_context_relevant(context, question, threshold=40):
+    """
+    Checks if the retrieved context is relevant to the question using fuzzy matching.
+    Returns True if relevant, False otherwise.
+    """
+    context_snippet = " ".join(context.split()[:100])  # Use only first 100 words for efficiency
+    relevance_score = fuzz.partial_ratio(context_snippet.lower(), question.lower())
+
+    return relevance_score >= threshold  # Only accept if score is above threshold
+
 def model_inference(retriever, question, llm):
+    # Validate question: Reject if too short or empty
+    if not question.strip() or len(question.strip()) < 5:
+        return "Answer:\n        The question is invalid or lacks sufficient detail."
+
     # Retrieve relevant documents
-    retrieved_docs = retriever.invoke(question)  # Ensure retriever returns documents
-    context = "\n\n".join([doc.page_content for doc in retrieved_docs])
+    retrieved_docs = retriever.invoke(question)
+
+    # If no relevant documents are found, reject the query
+    if not retrieved_docs:
+        return "Answer:\n        The answer is not found in the context provided.\nReferences: No references found."
+
+    # Extract context
+    context = "\n\n".join([doc.page_content for doc in retrieved_docs]).strip()
+
+    # **New: Check if the retrieved context is relevant to the question**
+    if not is_context_relevant(context, question):
+        return "Answer:\n        The answer is not found in the context provided.\nReferences: No references found."
 
     # Define prompt template
     prompt = PromptTemplate(
         input_variables=["context", "question"],
         template="""\
-        You are a helpful assistant. Answer the query accurately by focusing solely on the most relevant parts of the provided context.
-        1. Identify and use only the sections of the context directly related to the query, ignoring unrelated or extraneous information.
-        2. If the query cannot be explicitly answered using the relevant parts of the context, respond with: "The question is out of scope."
-        3. If you are unsure or do not know the answer, respond with: "I don't know.", don't try to make up an answer.
-        4. Do not add information, interpretations, or assumptions beyond what is explicitly stated in the context.
-        5. Ensure the response is concise, avoids redundancy, and directly addresses the query.
+        You are a highly accurate and reliable assistant. Follow these strict rules:
+
+        1. **Use Only the Provided Context** – Base your answer strictly on the given context. Do not infer or generate extra details.
+        2. **Reject Invalid or Unrelated Questions** – If the question does not relate to the context, respond with: "The answer is not found in the context provided."
+        3. **Ensure Context Relevance** – Answer only if the exact information exists in the context.
+        4. **No Hallucination** – Do not assume, summarize, or infer beyond what is explicitly stated.
+        5. **Concise and Relevant Answers** – Avoid redundancy. Provide only the most accurate response.
 
         Context:
         {context}
-        
+
         Query:
         {question}
-        
+
         Answer:
         """
     )
@@ -140,26 +168,21 @@ def model_inference(retriever, question, llm):
         RunnablePassthrough.assign(context=lambda _: context, question=lambda _: question)
         | prompt
         | llm
-        
     )
 
-    # Invoke the chain and extract the response
-    response = rag_chain.invoke({})
+    # Invoke the chain and get the response
+    response = rag_chain.invoke({}).strip()
 
-    # Extract the last occurrence of "Answer:" and everything after it
+    # Extract only the final "Answer:" section
     match = re.findall(r"Answer:\s*(.*)", response, re.DOTALL)
-    if match:
-        final_answer = "Answer:\n        " + match[-1].strip()  # Taking the last match
-    else:
-        final_answer = "Answer:\n        No valid answer found."
+    final_answer = "Answer:\n        " + match[-1].strip() if match else "Answer:\n        The answer is not found in the context provided."
 
-    # Generate references in the required format
-    references = [
-        f"Source: {doc.metadata['source']}, Page: {doc.metadata.get('page_label', 'Unknown')}" 
-        for doc in retrieved_docs
-    ]
+    # Extract only the first relevant reference
+    references = (
+        f"Source: {retrieved_docs[0].metadata['source']}, Page: {retrieved_docs[0].metadata.get('page_label', 'Unknown')}"
+        if retrieved_docs else "No references found."
+    )
 
-    # Return final structured output
     return f"{final_answer}\nReferences: {references}"
 
 # # Example usage: