app.py

import os
import re
import random
import requests
import streamlit as st
import spacy  # for additional NLP processing
from langchain_huggingface import HuggingFaceEndpoint
from langchain_core.prompts import PromptTemplate
from langchain_core.output_parsers import StrOutputParser
from transformers import pipeline

# Must be the first Streamlit command!
st.set_page_config(page_title="HAL - NASA ChatBot", page_icon="🚀")

# --- Helper to load spaCy model with fallback ---
def load_spacy_model():
    try:
        return spacy.load("en_core_web_sm")
    except OSError:
        st.warning("Downloading spaCy model en_core_web_sm... This may take a moment.")
        import subprocess
        subprocess.run(["python", "-m", "spacy", "download", "en_core_web_sm"], check=True)
        return spacy.load("en_core_web_sm")

nlp_spacy = load_spacy_model()

# --- 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",
    revision="714eb0f"
)

def get_llm_hf_inference(model_id=model_id, max_new_tokens=128, temperature=0.7):
    return HuggingFaceEndpoint(
        repo_id=model_id,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        token=os.getenv("HF_TOKEN"),
        task="text-generation"
    )

def get_nasa_apod():
    url = f"https://api.nasa.gov/planetary/apod?api_key={os.getenv('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.lower() or "space" in user_text.lower():
        return "nasa_info"
    return "general_query"

def extract_context(text):
    """
    Uses spaCy to extract named entities for additional context.
    """
    doc = nlp_spacy(text)
    entities = [ent.text for ent in doc.ents]
    return ", ".join(entities) if entities else ""

def generate_follow_up(user_text):
    """
    Generates two variant follow-up questions and randomly selects one.
    """
    prompt_text = (
        f"Based on the user's question: '{user_text}', generate two concise, friendly follow-up questions "
        "that invite further discussion. For example, one might be 'Would you like to know more about the six types of quarks?' "
        "and another '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()
    variants = re.split(r"\n|[;]+", output)
    cleaned = [v.strip(' "\'') for v in variants if v.strip()]
    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=512):
    """
    Generates HAL's detailed answer and a follow-up question.
    """
    sentiment = analyze_sentiment(user_text)
    action = predict_action(user_text)
    
    # Extract extra context
    context_info = extract_context(user_text)
    context_clause = f" The key topics here are: {context_info}." if context_info else ""
    
    # Extract style instructions if present.
    style_instruction = ""
    lower_text = user_text.lower()
    if "in the voice of" in lower_text or "speaking as" in lower_text:
        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}."
    
    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)
    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"
    
    style_clause = style_instruction if style_instruction else ""
    
    prompt = PromptTemplate.from_template(
        (
            "[INST] {system_message}\n\nCurrent Conversation:\n{chat_history}\n\n"
            "User: {user_text}.\n [/INST]\n"
            "AI: Please provide a detailed, in-depth answer in a friendly, conversational tone that thoroughly covers the topic."
            + style_clause + context_clause +
            "\nHAL:"
        )
    )
    
    # Debug: print the prompt for troubleshooting
    st.write("DEBUG: Prompt sent to model:")
    st.write(prompt.format(system_message=system_message, chat_history=filtered_history, user_text=user_text))
    
    chat = prompt | hf.bind(skip_prompt=True) | StrOutputParser(output_key='content')
    raw_output = chat.invoke(input=dict(system_message=system_message, user_text=user_text, chat_history=filtered_history))
    st.write("DEBUG: Raw model output:")
    st.write(raw_output)
    
    response = raw_output.split("HAL:")[-1].strip()
    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})
    
    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})
    
    st.write("DEBUG: Generated follow-up question:", 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!*")

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()

st.markdown("<div class='container'>", unsafe_allow_html=True)
for message in st.session_state.chat_history:
    if message["role"] == "user":
        st.markdown(f"<div class='user-msg'><strong>You:</strong> {message['content']}</div>", unsafe_allow_html=True)
    else:
        st.markdown(f"<div class='assistant-msg'><strong>HAL:</strong> {message['content']}</div>", unsafe_allow_html=True)
st.markdown("</div>", unsafe_allow_html=True)

user_input = st.chat_input("Type your message here...")

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