import os import re import requests import streamlit as st from langchain_huggingface import HuggingFaceEndpoint from langchain_core.prompts import PromptTemplate from langchain_core.output_parsers import StrOutputParser from transformers import pipeline # Use environment variables for keys HF_TOKEN = os.getenv("HF_TOKEN") if HF_TOKEN is None: raise ValueError("HF_TOKEN environment variable not set. Please set it in your Hugging Face Space settings.") NASA_API_KEY = os.getenv("NASA_API_KEY") if NASA_API_KEY is None: raise ValueError("NASA_API_KEY environment variable not set. Please set it in your Hugging Face Space settings.") # Set up Streamlit UI st.set_page_config(page_title="HAL - NASA ChatBot", page_icon="🚀") # --- Initialize Session State Variables --- if "chat_history" not in st.session_state: # The initial greeting is stored in chat_history. 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 # Tracks whether HAL has responded if "follow_up" not in st.session_state: st.session_state.follow_up = "" # Stores the follow-up question # --- Set Up Model & API Functions --- model_id = "mistralai/Mistral-7B-Instruct-v0.3" # Initialize sentiment analysis pipeline with explicit model specification sentiment_analyzer = pipeline( "sentiment-analysis", model="distilbert/distilbert-base-uncased-finetuned-sst-2-english", revision="714eb0f" ) def get_llm_hf_inference(model_id=model_id, max_new_tokens=128, temperature=0.7): # Specify task="text-generation" so that the endpoint uses the correct function. return HuggingFaceEndpoint( repo_id=model_id, max_new_tokens=max_new_tokens, temperature=temperature, token=HF_TOKEN, task="text-generation" ) def get_nasa_apod(): url = f"https://api.nasa.gov/planetary/apod?api_key={NASA_API_KEY}" response = requests.get(url) if response.status_code == 200: data = response.json() return data.get("url", ""), data.get("title", ""), data.get("explanation", "") else: return "", "NASA Data Unavailable", "I couldn't fetch data from NASA right now. Please try again later." def analyze_sentiment(user_text): result = sentiment_analyzer(user_text)[0] return result['label'] def predict_action(user_text): if "NASA" in user_text or "space" in user_text: return "nasa_info" return "general_query" def generate_follow_up(user_text): """ Generates a concise and conversational follow-up question related to the user's input. The prompt instructs the model to avoid meta commentary. """ prompt_text = ( f"Generate a concise, friendly follow-up question based on the user's question: '{user_text}'. " "Do not include meta instructions or commentary such as 'Never return just a statement.' " "For example, if the user asked about quarks, you might ask: " "'Would you like to know more about the six types of quarks, or is there another aspect of quantum physics you're curious about?'" ) hf = get_llm_hf_inference(max_new_tokens=64, temperature=0.8) follow_up = hf.invoke(input=prompt_text).strip() # Remove extraneous quotes if present. follow_up = follow_up.strip('\'"') # Optionally, remove any unwanted phrases (you can add more replacements if needed). follow_up = re.sub(r"Never return just a statement\.?", "", follow_up, flags=re.IGNORECASE).strip() # Ensure that something non-empty is returned. if not follow_up: follow_up = "Would you like to explore this topic further?" return follow_up def get_response(system_message, chat_history, user_text, max_new_tokens=256): """ Generates HAL's response in a friendly, conversational manner. Uses sentiment analysis to adjust tone when appropriate and always generates a follow-up question. If the user's input includes style instructions (e.g., 'in the voice of an astrophysicist'), the prompt instructs HAL to adapt accordingly. """ sentiment = analyze_sentiment(user_text) action = predict_action(user_text) # Check for style instructions in the user message. style_instruction = "" lower_text = user_text.lower() if "in the voice of" in lower_text or "speaking as" in lower_text: # Extract the style instruction (a simple heuristic: take the part after "in the voice of") match = re.search(r"(in the voice of|speaking as)(.*)", lower_text) if match: style_instruction = match.group(2).strip().capitalize() style_instruction = f" Please respond in the voice of {style_instruction}." # Handle NASA-related queries separately. if action == "nasa_info": nasa_url, nasa_title, nasa_explanation = get_nasa_apod() response = f"**{nasa_title}**\n\n{nasa_explanation}" chat_history.append({'role': 'user', 'content': user_text}) chat_history.append({'role': 'assistant', 'content': response}) follow_up = generate_follow_up(user_text) chat_history.append({'role': 'assistant', 'content': follow_up}) return response, follow_up, chat_history, nasa_url hf = get_llm_hf_inference(max_new_tokens=max_new_tokens, temperature=0.9) # Build a filtered conversation history excluding the initial greeting. filtered_history = "" for message in chat_history: if message["role"] == "assistant" and message["content"].strip() == "Hello! How can I assist you today?": continue filtered_history += f"{message['role']}: {message['content']}\n" # Add style instruction to the prompt if applicable. style_clause = "" if style_instruction: style_clause = style_instruction 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 any previous greetings." " Keep your response friendly and conversational, 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)) response = response.split("HAL:")[-1].strip() chat_history.append({'role': 'user', 'content': user_text}) chat_history.append({'role': 'assistant', 'content': response}) # Only override with an empathetic response for negative sentiment if the input is not a direct question. if sentiment == "NEGATIVE" and not user_text.strip().endswith("?"): response = "I'm sorry you're feeling this way. I'm here to help. What can I do to assist you further?" chat_history[-1]['content'] = response follow_up = generate_follow_up(user_text) chat_history.append({'role': 'assistant', 'content': follow_up}) return response, follow_up, chat_history, None # --- Chat UI --- st.title("🚀 HAL - Your NASA AI Assistant") st.markdown("🌌 *Ask me about space, NASA, and beyond!*") # Sidebar: Reset Chat if st.sidebar.button("Reset Chat"): st.session_state.chat_history = [{"role": "assistant", "content": "Hello! How can I assist you today?"}] st.session_state.response_ready = False st.session_state.follow_up = "" st.experimental_rerun() # Custom Chat Styling st.markdown(""" """, unsafe_allow_html=True) # --- Single Input Box for Both Initial and Follow-Up Messages --- user_input = st.chat_input("Type your message here...") # Only ONE chat_input() if user_input: response, follow_up, st.session_state.chat_history, image_url = get_response( system_message="You are a helpful AI assistant.", 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 # Render the entire chat history. st.markdown("
", unsafe_allow_html=True) for message in st.session_state.chat_history: if message["role"] == "user": st.markdown(f"
You: {message['content']}
", unsafe_allow_html=True) else: st.markdown(f"
HAL: {message['content']}
", unsafe_allow_html=True) st.markdown("
", unsafe_allow_html=True)