Update app.py

app.py

@@ -23,16 +23,13 @@ st.set_page_config(page_title="HAL - NASA ChatBot", page_icon="🚀")
 # --- Initialize Session State Variables ---
 if "chat_history" not in st.session_state:
     st.session_state.chat_history = [{"role": "assistant", "content": "Hello! How can I assist you today?"}]
-
 if "response_ready" not in st.session_state:
     st.session_state.response_ready = False
-
 if "follow_up" not in st.session_state:
     st.session_state.follow_up = ""
 
 # --- Set Up Model & API Functions ---
 model_id = "mistralai/Mistral-7B-Instruct-v0.3"
-
 sentiment_analyzer = pipeline(
     "sentiment-analysis",
     model="distilbert/distilbert-base-uncased-finetuned-sst-2-english",
@@ -68,35 +65,32 @@ def predict_action(user_text):
 
 def generate_follow_up(user_text):
     """
-    Generates varied follow-up questions for the given user input.
-    The prompt instructs the LLM to produce two variants, and one is selected randomly.
+    Generates two variant follow-up questions and randomly selects one.
+    It also cleans up any unwanted quotation marks or extra meta commentary.
     """
     prompt_text = (
         f"Based on the user's question: '{user_text}', generate two concise, friendly follow-up questions "
-        "that are relevant to the topic. One should ask something like, "
-        "'Would you like to know more about the six types of quarks?' and the other should ask, "
-        "'Would you like to explore something else?' Do not include any extra commentary or meta instructions."
+        "that invite further discussion. For example, one might be 'Would you like to know more about the six types of quarks?' "
+        "and another might be 'Would you like to explore another aspect of quantum physics?' Do not include extra commentary."
     )
     hf = get_llm_hf_inference(max_new_tokens=80, temperature=0.9)
     output = hf.invoke(input=prompt_text).strip()
-    # Split the output into separate lines if the model returns multiple variants.
     variants = re.split(r"\n|[;]+", output)
-    # Clean up any extraneous quotes or unwanted text.
     cleaned = [v.strip(' "\'') for v in variants if v.strip()]
-    # If no valid variants are found, provide a default fallback.
     if not cleaned:
         cleaned = ["Would you like to explore this topic further?"]
     return random.choice(cleaned)
 
 def get_response(system_message, chat_history, user_text, max_new_tokens=256):
     """
-    Generates HAL's response with a friendly, conversational tone.
-    Incorporates sentiment analysis and always generates a follow-up question with variation.
+    Generates HAL's answer with depth and a follow-up question.
+    The prompt instructs the model to provide a detailed explanation and then generate a follow-up.
+    If the answer comes back empty, a fallback answer is used.
     """
     sentiment = analyze_sentiment(user_text)
     action = predict_action(user_text)
     
-    # Check for style instructions in the user's text (e.g., "in the voice of an astrophysicist")
+    # Extract style instruction if present
     style_instruction = ""
     lower_text = user_text.lower()
     if "in the voice of" in lower_text or "speaking as" in lower_text:
@@ -115,7 +109,6 @@ def get_response(system_message, chat_history, user_text, max_new_tokens=256):
         return response, follow_up, chat_history, nasa_url
     
     hf = get_llm_hf_inference(max_new_tokens=max_new_tokens, temperature=0.9)
-    
     filtered_history = ""
     for message in chat_history:
         if message["role"] == "assistant" and message["content"].strip() == "Hello! How can I assist you today?":
@@ -124,21 +117,27 @@ def get_response(system_message, chat_history, user_text, max_new_tokens=256):
     
     style_clause = style_instruction if style_instruction else ""
     
+    # Instruct the model to generate a detailed, in-depth answer.
     prompt = PromptTemplate.from_template(
         (
             "[INST] {system_message}\n\nCurrent Conversation:\n{chat_history}\n\n"
             "User: {user_text}.\n [/INST]\n"
-            "AI: Please answer the user's question without repeating previous greetings. "
-            "Keep your response friendly and conversational, starting with a phrase like "
-            "'Certainly!', 'Of course!', or 'Great question!'." + style_clause +
+            "AI: Please provide a detailed explanation in depth. "
+            "Ensure your response covers the topic thoroughly and is written in a friendly, conversational style, "
+            "starting with a phrase like 'Certainly!', 'Of course!', or 'Great question!'." + style_clause +
             "\nHAL:"
         )
     )
     
     chat = prompt | hf.bind(skip_prompt=True) | StrOutputParser(output_key='content')
     response = chat.invoke(input=dict(system_message=system_message, user_text=user_text, chat_history=filtered_history))
+    # Remove any extra markers if present.
     response = response.split("HAL:")[-1].strip()
     
+    # Fallback in case the generated answer is empty
+    if not response:
+        response = "Certainly, here is an in-depth explanation: [Fallback explanation]."
+    
     chat_history.append({'role': 'user', 'content': user_text})
     chat_history.append({'role': 'assistant', 'content': response})
     
@@ -200,10 +199,8 @@ if user_input:
         user_text=user_input,
         chat_history=st.session_state.chat_history
     )
-    
     if image_url:
         st.image(image_url, caption="NASA Image of the Day")
-    
     st.session_state.follow_up = follow_up
     st.session_state.response_ready = True