app.py

import os
import re
import random
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
from langdetect import detect

# 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:
    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=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 is_apod_query(user_text):
    """
    Checks if the user's question contains keywords indicating they are asking for
    the Astronomy Picture of the Day.
    """
    keywords = ["apod", "image", "picture", "photo", "astronomy picture"]
    return any(keyword in user_text.lower() for keyword in keywords)

def predict_action(user_text):
    """
    Determines the type of response for the user's query.
    If the query contains "NASA" or "space" and also indicates an APOD query,
    it returns "nasa_info". Otherwise, it returns "general_query", even if "NASA" is present.
    """
    user_text_lower = user_text.lower()
    if "nasa" in user_text_lower or "space" in user_text_lower:
        if is_apod_query(user_text):
            return "nasa_info"
    return "general_query"

def generate_follow_up(user_text):
    """
    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 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 ."
        "Answer exclusively in English, and 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)

from langdetect import detect

def ensure_english(text):
    """Check if the model accidentally generated a non-English response."""
    try:
        detected_lang = detect(text)
        if detected_lang != "en":
            return "⚠️ Sorry, I only respond in English. Can you rephrase your question?"
    except:
        return "⚠️ Language detection failed. Please ask your question again."
    return text

def get_response(system_message, chat_history, user_text, max_new_tokens=512):
    """
    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)

    # 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:
        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 ""
    
    # 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 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!'."
        "🚨 IMPORTANT: Answer exclusively in **English only**. Do not generate responses in any other language."
        "\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()

    # 🚨 Ensure the response is in English
    response = ensure_english(response)


    # 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})

    # 🔍 FIX: Only override if strongly negative and NOT a 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 = None  # 🚨 Don't generate follow-up if negative sentiment triggers a different message
    else:
        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!*")

#Reset Button
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()

#Style and Appearance
st.markdown("""
    <style>
    .user-msg {
        background-color: #696969;
        color: white;
        padding: 10px;
        border-radius: 10px;
        margin-bottom: 5px;
        width: fit-content;
        max-width: 80%;
    }
    .assistant-msg {
        background-color: #333333;
        color: white;
        padding: 10px;
        border-radius: 10px;
        margin-bottom: 5px;
        width: fit-content;
        max-width: 80%;
    }
    .container {
        display: flex;
        flex-direction: column;
        align-items: flex-start;
    }
    @media (max-width: 600px) {
        .user-msg, .assistant-msg { font-size: 16px; max-width: 100%; }
    }
    </style>
""", 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

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)