src/streamlit_app.py

import streamlit as st
import tensorflow as tf
import numpy as np
import nltk
import os
from nltk.tokenize import sent_tokenize
from transformers import DistilBertTokenizerFast, TFDistilBertForSequenceClassification

# 📁 Hugging Face cache dir
os.environ["TRANSFORMERS_CACHE"] = "/tmp/huggingface"

# 📥 Download NLTK punkt tokenizer
nltk_data_path = "/tmp/nltk_data"
nltk.download("punkt_tab", download_dir=nltk_data_path)
nltk.data.path.append(nltk_data_path)

# ✅ Cache the model/tokenizer
@st.cache_resource
def load_model_and_tokenizer():
    tokenizer = DistilBertTokenizerFast.from_pretrained(
        "distilbert-base-uncased", cache_dir="/tmp/huggingface"
    )
    model = TFDistilBertForSequenceClassification.from_pretrained(
        "sundaram07/distilbert-sentence-classifier", cache_dir="/tmp/huggingface"
    )
    return tokenizer, model

tokenizer, model = load_model_and_tokenizer()

# 🔮 Predict sentence AI probability
def predict_sentence_ai_probability(sentence):
    inputs = tokenizer(sentence, return_tensors="tf", truncation=True, padding=True)
    outputs = model(inputs)
    logits = outputs.logits
    prob_ai = tf.sigmoid(logits)[0][0].numpy()
    return prob_ai

# 📊 Analyze text
def predict_ai_generated_percentage(text, threshold=0.15):
    text = text.strip()
    sentences = sent_tokenize(text)
    if len(sentences) == 0:
        return 0.0, []

    ai_sentence_count = 0
    results = []

    for sentence in sentences:
        prob = predict_sentence_ai_probability(sentence)
        is_ai = prob <= threshold
        results.append((sentence, prob, is_ai))
        if is_ai:
            ai_sentence_count += 1

    ai_percentage = (ai_sentence_count / len(sentences)) * 100
    return ai_percentage, results

# 🖥️ Streamlit UI
st.set_page_config(page_title="AI Detector", layout="wide")
st.title("🧠 AI Content Detector")
st.markdown("This app detects the percentage of **AI-generated content** using sentence-level analysis with DistilBERT.")

# 📋 Text input
user_input = st.text_area("📋 Paste your text below to check for AI-generated sentences:", height=300)

# 🔍 Analyze button logic
if st.button("🔍 Analyze"):
    # Clear previous session results
    st.session_state.analysis_done = False
    st.session_state.analysis_results = None
    st.session_state.ai_percentage = None

    if not user_input.strip():
        st.warning("⚠️ Please enter some text.")
    else:
        # Perform analysis
        ai_percentage, analysis_results = predict_ai_generated_percentage(user_input)

        if len(analysis_results) == 0:
            st.warning("⚠️ Not enough valid sentences to analyze.")
        else:
            st.session_state.analysis_done = True
            st.session_state.analysis_results = analysis_results
            st.session_state.ai_percentage = ai_percentage

# 📤 Show results
if st.session_state.get("analysis_done", False):
    st.subheader("🔍 Sentence-level Analysis")
    for i, (sentence, prob, is_ai) in enumerate(st.session_state.analysis_results, start=1):
        label = "🟢 Human" if not is_ai else "🔴 AI"
        st.markdown(f"**{i}.** _{sentence}_\n\n → {label}")

    st.subheader("📊 Final Result")
    st.success(f"Estimated **AI-generated content**: **{st.session_state.ai_percentage:.2f}%**")