app.py

import os
import json
import ast
import streamlit as st
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import re
import math
import logging

st.set_page_config(
    page_title="AI Article Detection by DEJAN",
    page_icon="🧠",
    layout="wide"
)

# --- Load heuristic weights from environment secrets, with JSON→Python fallback ---
@st.cache_resource
def load_heuristic_weights():
    def _load(env_key):
        raw = os.environ[env_key]
        try:
            return json.loads(raw)
        except json.JSONDecodeError:
            return ast.literal_eval(raw)
    ai = _load("AI_WEIGHTS_JSON")
    og = _load("OG_WEIGHTS_JSON")
    return ai, og

AI_WEIGHTS, OG_WEIGHTS = load_heuristic_weights()
SIGMOID_K = 0.5

def tokenize(text):
    return re.findall(r'\b[a-z]{2,}\b', text.lower())

def classify_text_likelihood(text: str) -> float:
    tokens = tokenize(text)
    if not tokens:
        return 0.5
    ai_score = og_score = matched = 0
    for t in tokens:
        aw = AI_WEIGHTS.get(t, 0)
        ow = OG_WEIGHTS.get(t, 0)
        if aw or ow:
            matched += 1
            ai_score += aw
            og_score += ow
    if matched == 0:
        return 0.5
    net = ai_score - og_score
    return 1 / (1 + math.exp(-SIGMOID_K * net))

def highlight_heuristic_words(text: str) -> str:
    parts = re.split(r'(\b[a-z]{2,}\b)', text)
    out = []
    for part in parts:
        lower = part.lower()
        if lower in AI_WEIGHTS:
            out.append(
                f"<span style='text-decoration: underline; "
                f"text-decoration-color: darkred; text-decoration-thickness: 2px;'>"
                f"{part}</span>"
            )
        elif lower in OG_WEIGHTS:
            out.append(
                f"<span style='text-decoration: underline; "
                f"text-decoration-color: darkgreen; text-decoration-thickness: 2px;'>"
                f"{part}</span>"
            )
        else:
            out.append(part)
    return ''.join(out)

# --- Logging & Streamlit setup ---
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)



st.markdown("""
<link href="https://fonts.googleapis.com/css2?family=Roboto&display=swap" rel="stylesheet">
<style>
    html, body, [class*="css"] {
        font-family: 'Roboto', sans-serif;
    }
</style>
""", unsafe_allow_html=True)

@st.cache_resource
def load_model_and_tokenizer(model_name):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    dtype = torch.bfloat16 if (device.type=="cuda" and torch.cuda.is_bf16_supported()) else torch.float32
    model = AutoModelForSequenceClassification.from_pretrained(model_name, torch_dtype=dtype)
    model.to(device).eval()
    return tokenizer, model, device

MODEL_NAME = "dejanseo/ai-detection-small"
try:
    tokenizer, model, device = load_model_and_tokenizer(MODEL_NAME)
except Exception as e:
    st.error(f"Error loading model: {e}")
    logger.error(f"Failed to load model: {e}", exc_info=True)
    st.stop()

def sent_tokenize(text):
    return [s for s in re.split(r'(?<=[\.!?])\s+', text.strip()) if s]

st.title("AI Article Detection")

text = st.text_area("Enter text to classify", height=200, placeholder="Paste your text here…")

if st.button("Classify", type="primary"):
    if not text.strip():
        st.warning("Please enter some text.")
    else:
        with st.spinner("Analyzing…"):
            sentences = sent_tokenize(text)
            if not sentences:
                st.warning("No sentences detected.")
                st.stop()

            inputs = tokenizer(
                sentences,
                return_tensors="pt",
                padding=True,
                truncation=True,
                max_length=model.config.max_position_embeddings
            ).to(device)

            with torch.no_grad():
                logits = model(**inputs).logits
                probs = F.softmax(logits, dim=-1).cpu()
                preds = torch.argmax(probs, dim=-1).cpu()

            chunks = []
            for i, s in enumerate(sentences):
                inner = highlight_heuristic_words(s)
                p = preds[i].item()
                r, g = (255, 0) if p == 0 else (0, 255)
                conf = probs[i, p].item()
                alpha = conf
                span = (
                    f"<span style='background-color: rgba({r},{g},0,{alpha:.2f}); "
                    f"padding:2px; margin:0 2px; border-radius:3px;'>{inner}</span>"
                )
                chunks.append(span)
            st.markdown("".join(chunks), unsafe_allow_html=True)

            avg = torch.mean(probs, dim=0)
            model_ai = avg[0].item()
            heuristic_ai = classify_text_likelihood(text)
            combined = min(model_ai + heuristic_ai, 1.0)

            st.subheader(f"🤖 Model AI Likelihood: {model_ai*100:.1f}%")
            st.subheader(f"🛠️ Heuristic AI Likelihood: {heuristic_ai*100:.1f}%")
            st.subheader(f"⚖️ Combined AI Likelihood: {combined*100:.1f}%")