app.py

import os
import time
from datetime import datetime
import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# -- SETUP --
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"

@st.cache_resource
def load_model():
    model_id = "tiiuae/falcon-rw-1b"
    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
    return tokenizer, model

tokenizer, model = load_model()

if "history" not in st.session_state:
    st.session_state.history = []
    st.session_state.summary = ""

# -- SAFETY --
TRIGGER_PHRASES = ["kill myself", "end it all", "suicide", "not worth living", "can't go on"]
def is_high_risk(text):
    return any(phrase in text.lower() for phrase in TRIGGER_PHRASES)

# -- GENERATE FUNCTION --
def generate_response(prompt, max_new_tokens=120, temperature=0.7):
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True)
    outputs = model.generate(
        **inputs,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        temperature=temperature,
        pad_token_id=tokenizer.eos_token_id
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True).strip()

# -- STYLING --
st.markdown("""
    <style>
    body {
        background-color: #111827;
        color: #f3f4f6;
    }
    .stTextInput > div > div > input {
        color: #f3f4f6;
    }
    </style>
""", unsafe_allow_html=True)

# -- HEADER --
st.title("🧠 TARS.help")
st.markdown("### A minimal AI that listens, reflects, and replies.")
st.markdown(f"🗓️ {datetime.now().strftime('%B %d, %Y')} | {len(st.session_state.history)//2} exchanges")

# -- USER INPUT --
user_input = st.text_input("How are you feeling today?", placeholder="Start typing...")

# -- CHAT FLOW --
if user_input:
    context = "\n".join([f"{s}: {m}" for s, m, _ in st.session_state.history[-4:]])
    with st.spinner("TARS is reflecting..."):
        time.sleep(0.8)
        if is_high_risk(user_input):
            response = "I'm really sorry you're feeling this way. You're not alone — please talk to someone you trust or a mental health professional. 💙"
        else:
            prompt = f"You are an empathetic AI. Here's the recent conversation:\n{context}\nUser: {user_input}\nAI:"
            response = generate_response(prompt, max_new_tokens=100)
            response = response.split("AI:")[-1].strip()
    timestamp = datetime.now().strftime("%H:%M")
    st.session_state.history.append(("🧍 You", user_input, timestamp))
    st.session_state.history.append(("🤖 TARS", response, timestamp))

# -- DISPLAY CHAT --
st.markdown("## 🗨️ Session")
for speaker, msg, time in st.session_state.history:
    st.markdown(f"**{speaker} [{time}]:** {msg}")

# -- SUMMARY --
if st.button("🧾 Generate Session Summary"):
    convo = "\n".join([f"{s}: {m}" for s, m, _ in st.session_state.history])
    summary_prompt = f"Summarize this conversation in 2-3 sentences:\n{convo}\nSummary:"
    try:
        summary = generate_response(summary_prompt, max_new_tokens=150, temperature=0.5)
        st.session_state.summary = summary.split("Summary:")[-1].strip()
    except Exception as e:
        st.error("❌ Failed to generate summary.")
        st.exception(e)

# -- DISPLAY SUMMARY --
if st.session_state.summary:
    st.markdown("### 🧠 Session Note")
    st.markdown(st.session_state.summary)
    st.download_button("📥 Download Summary", st.session_state.summary, file_name="tars_session.txt")

# -- FOOTER --
st.markdown("---")
st.caption("TARS is not a therapist but a quiet assistant that reflects with you.")