app.py

import os
import re
import json
import time
import datetime
import numpy as np
import joblib
import pandas as pd
import traceback
import logging
from collections import defaultdict
from flask import Flask, request, jsonify
from flask_cors import CORS
from werkzeug.security import generate_password_hash, check_password_hash
import requests
import jwt
from feature_extraction import FeatureExtractor
from source_reputation import ReputationEngine
from init_db import init_db
from db import get_conn, PH, IntegrityError

# ─── Logging ────────────────────────────────────────────────────────────────
logging.basicConfig(level=logging.INFO,
                    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# ─── App Setup ───────────────────────────────────────────────────────────────
app = Flask(__name__)

# Allow requests from the Vercel frontend (set FRONTEND_ORIGIN env var in production)
_frontend_origin = os.environ.get('FRONTEND_ORIGIN', '*')
CORS(app, origins=_frontend_origin, supports_credentials=True)

SECRET_KEY = os.environ.get('TRUTH_SECRET_KEY', 'truth-detector-jwt-secret-2025')

# ─── Rate Limiter (in-memory, per IP) ────────────────────────────────────────
_login_attempts = defaultdict(list)   # ip -> [timestamp, ...]
RATE_LIMIT_MAX   = 5     # max attempts
RATE_LIMIT_WINDOW = 300  # seconds (5 minutes)


def _check_rate_limit(ip: str) -> bool:
    """Returns True if the IP is allowed, False if rate-limited."""
    now = time.time()
    attempts = [t for t in _login_attempts[ip] if now - t < RATE_LIMIT_WINDOW]
    _login_attempts[ip] = attempts
    if len(attempts) >= RATE_LIMIT_MAX:
        return False
    _login_attempts[ip].append(now)
    return True


# ─── Input Validators ────────────────────────────────────────────────────────
USERNAME_RE = re.compile(r'^[A-Za-z0-9_]{3,30}$')
EMAIL_RE    = re.compile(r'^[^\s@]+@[^\s@]+\.[^\s@]+$')


def validate_username(u: str):
    if not u:
        return "Username is required."
    if not USERNAME_RE.match(u):
        return "Username must be 3–30 characters: letters, digits, or underscore only."
    return None


def validate_password(p: str):
    if not p:
        return "Password is required."
    if len(p) < 8:
        return "Password must be at least 8 characters."
    if not re.search(r'[A-Za-z]', p):
        return "Password must contain at least one letter."
    if not re.search(r'[0-9]', p):
        return "Password must contain at least one number."
    return None


def validate_email(e: str):
    if e and not EMAIL_RE.match(e):
        return "Invalid email format."
    return None


# ─── JWT Helpers ──────────────────────────────────────────────────────────────
def _issue_token(username: str) -> str:
    payload = {
        'sub': username,
        'iat': datetime.datetime.utcnow(),
        'exp': datetime.datetime.utcnow() + datetime.timedelta(hours=24)
    }
    return jwt.encode(payload, SECRET_KEY, algorithm='HS256')


def _decode_token(token: str):
    """Returns username on success, or raises jwt.PyJWTError."""
    payload = jwt.decode(token, SECRET_KEY, algorithms=['HS256'])
    return payload['sub']


def _get_request_token() -> str | None:
    auth = request.headers.get('Authorization', '')
    if auth.startswith('Bearer '):
        return auth[7:]
    # Fallback: let legacy localStorage pass username directly (for predict route)
    return request.headers.get('X-Username')


# ─── Global Models & Extractor ────────────────────────────────────────────────
models = {}
metrics = {}
thresholds = {}

# ─── Model & Asset Loading ──────────────────────────────────────────────────
extractor = FeatureExtractor()
reputation_engine = ReputationEngine()


def load_models():
    global models, metrics, thresholds
    base_dir = os.path.dirname(os.path.abspath(__file__))
    models_dir = os.path.join(base_dir, 'models')

    try:
        model_names = ['nb', 'lr', 'svm', 'rf', 'dl', 'ensemble']
        for name in model_names:
            path = os.path.join(models_dir, f'{name}_model.pkl')
            if os.path.exists(path):
                models[name] = joblib.load(path)
                logger.info(f"Loaded model: {name}")

        metrics_path = os.path.join(models_dir, 'metrics.json')
        if os.path.exists(metrics_path):
            with open(metrics_path, 'r') as f:
                metrics = json.load(f)

        thresholds_path = os.path.join(models_dir, 'thresholds.json')
        if os.path.exists(thresholds_path):
            with open(thresholds_path, 'r') as f:
                thresholds = json.load(f)
                logger.info(f"Loaded thresholds: {list(thresholds.keys())}")
        else:
            logger.warning("No thresholds.json found, using default 0.5")
    except Exception as e:
        logger.error(f"Error loading models: {e}")


# ─── Global Error Handler ────────────────────────────────────────────────────
@app.errorhandler(Exception)
def handle_exception(e):
    if hasattr(e, 'code') and isinstance(e.code, int):
        return jsonify({'error': str(e)}), e.code
    logger.error(f"Unhandled Exception: {traceback.format_exc()}")
    return jsonify({'error': 'Internal Server Error', 'message': str(e)}), 500


# ─── HEALTH ROUTE ─────────────────────────────────────────────────────────────
@app.route('/api/health', methods=['GET'])
@app.route('/health', methods=['GET'])
def health():
    return jsonify({
        'status': 'ok',
        'models_loaded': list(models.keys()),
        'extractor_ready': extractor is not None,
        'timestamp': datetime.datetime.utcnow().isoformat()
    })


# ─── PREDICTION ROUTE ─────────────────────────────────────────────────────────
@app.route('/api/predict', methods=['POST'])
@app.route('/predict', methods=['POST'])
def predict():
    try:
        data = request.json
        if not data:
            return jsonify({'error': 'No data provided'}), 400

        statement  = data.get('statement', '').strip()
        model_type = data.get('model', 'ensemble')

        if not statement:
            return jsonify({'error': 'No statement provided'}), 400
        if len(statement) < 5:
            return jsonify({'error': 'Statement is too short (min 5 characters)'}), 400

        logger.info(f"Prediction request [{model_type}]: {statement[:60]}...")

        if extractor is None:
            return jsonify({'error': 'Feature extractor failed to initialize'}), 500

        # ── §3.3 Hybrid Prediction ────────────────────────────────────────────
        input_data = {**raw_features, 'statement': statement}
        input_df   = pd.DataFrame([input_data])

        if model_type == 'ensemble' and 'ensemble' not in models and len(models) > 0:
            probs = []
            for m_name, m_obj in models.items():
                try:
                    p = m_obj.predict_proba(input_df)[0].tolist()[1]
                    probs.append(p)
                except Exception:
                    pass
            real_prob = float(np.mean(probs)) if probs else 0.5
            prob = [1.0 - real_prob, real_prob]
        else:
            model = models.get(model_type) or (list(models.values())[0] if models else None)
            if not model:
                return jsonify({'error': 'No model available.'}), 500
            try:
                prob = model.predict_proba(input_df)[0].tolist()
            except Exception:
                prob = [0.5, 0.5]
            real_prob = prob[1]  # P(Real)

        # Calibrated threshold
        t_data    = thresholds.get(model_type, {})
        threshold = t_data.get('threshold', 0.5) if t_data else 0.5
        prediction = 1 if real_prob >= threshold else 0

        # Confidence: distance from decision boundary
        confidence = real_prob * 100 if prediction == 1 else (1 - real_prob) * 100

        # ── §3.3 Logic Guardrail (fixed — bounded probability shift only) ─────
        is_official   = raw_features.get('official_marker', 0) > 0
        formal_cadence = raw_features.get('formal_cadence', 0)
        sensationalism_score = raw_features.get('sensationalism_score', 0)
        sensational_hit_count = raw_features.get('sensational_hit_count', 0)

        # Credibility boost ONLY for verified official sources with strong formal cadence
        if is_official and formal_cadence > 0.08:
            logger.info("Guardrail: Official source detected — applying credibility boost.")
            # Safe bounded shift: at most +5% to real_prob
            boost = min(0.05, (1.0 - real_prob) * 0.15)
            real_prob = real_prob + boost
            if real_prob >= threshold:
                prediction = 1
            confidence = real_prob * 100 if prediction == 1 else (1 - real_prob) * 100

        # Sensationalism penalty — reduce real_prob for fake-news signature patterns
        if sensational_hit_count > 0:
            penalty = min(0.50, sensational_hit_count * 0.15)
            real_prob = max(0.0, real_prob - penalty)
            if real_prob < threshold:
                prediction = 0
            confidence = real_prob * 100 if prediction == 1 else (1 - real_prob) * 100
            logger.info(f"Sensationalism penalty: -{penalty:.2f} (hits={sensational_hit_count})")

        # ── Final label ───────────────────────────────────────────────────────
        final_label = 'Real' if prediction == 1 else 'Fake'

        if confidence < 50.1:
            final_label = 'Uncertain'
            logger.info("Confidence below 50.1% — labeling as 'Uncertain'")

        logger.info(f"Result: {final_label} | Confidence: {confidence:.2f}%")

        # ── §3 Paper Taxonomy Breakdown ───────────────────────────────────────
        paper_taxonomy = {
            # §3.2.1 News Content Features
            'content_features': {
                'lexical_density':    round(raw_features.get('lexical_density', 0), 4),
                'sentiment_score':    round(raw_features.get('sentiment_score', 0), 4),
                'complexity_score':   round(raw_features.get('complexity_score', 0), 4),
                'subjectivity_score': round(raw_features.get('subjectivity_score', 0), 4),
                'emotional_intensity': round(raw_features.get('emotional_intensity', 0), 4),
            },
            # §3.2.2 Social Context Features
            'social_context': {
                'speaker_reliability': round(raw_features.get('speaker_reliability', 0.5), 4),
                'false_history_ratio': round(raw_features.get('false_history_ratio', 0), 4),
                'true_history_ratio':  round(raw_features.get('true_history_ratio', 0), 4),
                'history_volume':      round(raw_features.get('history_volume', 0), 4),
                'is_republican':       int(raw_features.get('is_republican', 0)),
                'is_democrat':         int(raw_features.get('is_democrat', 0)),
            },
            # §3.3 Knowledge-Guided Signals
            'knowledge_signals': {
                'official_marker':    int(raw_features.get('official_marker', 0)),
                'formal_cadence':     round(raw_features.get('formal_cadence', 0), 4),
                'attribution_ratio':  round(raw_features.get('attribution_ratio', 0), 4),
                'has_source_citation': int(raw_features.get('has_source_citation', 0)),
                'entity_density':     round(raw_features.get('entity_density', 0), 4),
            },
            # §3.3.2 Deep Module Signals
            'deep_module': {
                'hedge_ratio':         round(raw_features.get('hedge_ratio', 0), 4),
                'certainty_ratio':     round(raw_features.get('certainty_ratio', 0), 4),
                'negation_ratio':      round(raw_features.get('negation_ratio', 0), 4),
                'caps_word_ratio':     round(raw_features.get('caps_word_ratio', 0), 4),
                'number_density':      round(raw_features.get('number_density', 0), 4),
            },
            # Deception & Sensationalism Signals
            'deception_signals': {
                'sensationalism_score': round(raw_features.get('sensationalism_score', 0), 4),
                'conspiracy_score':    round(raw_features.get('conspiracy_score', 0), 4),
                'health_misinfo_score': round(raw_features.get('health_misinfo_score', 0), 4),
                'sensational_hit_count': int(raw_features.get('sensational_hit_count', 0)),
                'absolutist_ratio':    round(raw_features.get('absolutist_ratio', 0), 4),
            }
        }

        # Legacy taxonomy breakdown kept for backward compatibility
        taxonomy_breakdown = {
            'lexical_density':        raw_features.get('lexical_density', 0),
            'syntactic_noun_ratio':   raw_features.get('noun_ratio', 0),
            'style_capital_ratio':    raw_features.get('capital_ratio', 0),
            'social_reliability_proxy': raw_features.get('speaker_reliability', 0.5)
        }

        # Source Reputation Analysis
        speaker = data.get('speaker')
        context_str = data.get('context')
        source_rep = reputation_engine.analyze_source(statement, speaker=speaker, context=context_str)

        return jsonify({
            'prediction':         final_label,
            'confidence':         round(confidence, 2),
            'probabilities':      {'fake': round(prob[0], 4), 'real': round(prob[1], 4)},
            'taxonomy_breakdown': taxonomy_breakdown,
            'paper_taxonomy':     paper_taxonomy,
            'model_used':         model_type,
            'source_reputation':  source_rep
        })

    except Exception as e:
        logger.error(f"Predict Error: {traceback.format_exc()}")
        return jsonify({'error': 'Prediction failed', 'details': str(e)}), 500


# ─── METRICS ROUTE ────────────────────────────────────────────────────────────
@app.route('/api/metrics', methods=['GET'])
@app.route('/metrics', methods=['GET'])
def get_metrics():
    return jsonify(metrics)


# ─── MODEL RECOMMENDATION ROUTE ───────────────────────────────────────────────
@app.route('/api/recommend-model', methods=['POST'])
@app.route('/recommend-model', methods=['POST'])
def recommend_model():
    """Analyze statement and recommend the best model (§3 taxonomy heuristic)."""
    try:
        data = request.json
        statement = data.get('statement', '').strip()
        if not statement or len(statement) < 10:
            return jsonify({'recommended': 'ensemble', 'reason': 'Enter more text.',
                            'scores': {}, 'reasons': {}, 'model_info': {}})
        if extractor is None:
            return jsonify({'recommended': 'ensemble', 'reason': 'Extractor unavailable.',
                            'scores': {}, 'reasons': {}, 'model_info': {}})

        metadata = {'party': 'independent', 'barely_true_counts': 0, 'false_counts': 0,
                    'half_true_counts': 0, 'mostly_true_counts': 0, 'pants_on_fire_counts': 0}
        feats = extractor.get_combined_features(statement, metadata)

        scores, reasons = {}, {}
        wc = feats.get('total_words', 0)
        sx = abs(feats.get('sentiment_score', 0))
        em = feats.get('emotional_intensity', 0)
        fm = feats.get('formal_cadence', 0)
        cx = feats.get('complexity_score', 0)
        nr = feats.get('noun_ratio', 0)
        nd = feats.get('number_density', 0)
        hs = feats.get('has_source_citation', 0)
        hg = feats.get('hedge_ratio', 0)
        cp = feats.get('caps_word_ratio', 0)
        sl = feats.get('avg_sentence_length', 0)

        # NB: emotional, short, keyword-heavy text
        s, r = 0.5, []
        if wc < 30:       s += 0.2; r.append('short text')
        if em > 0.02:     s += 0.2; r.append('high emotional language')
        if sx > 0.5:      s += 0.15; r.append('strong sentiment')
        if cp > 0.05:     s += 0.1; r.append('urgency markers')
        scores['nb'] = round(s, 3)
        reasons['nb'] = 'Strong for ' + ', '.join(r) if r else 'General keyword analysis'

        # LR: formal, noun-heavy, source-cited - HIGHLY RECOMMENDED FOR COMPLEXITY
        s, r = 0.5, []
        if nr > 0.15:     s += 0.2; r.append('high noun density')
        if fm > 0.02:     s += 0.15; r.append('formal structure')
        if hs > 0:        s += 0.15; r.append('source citations')
        if nd > 0.05:     s += 0.1; r.append('statistical content')
        if cx < 40 or sl > 15: s += 0.3; r.append('structural complexity')
        if 40 < cx < 70:  s += 0.1; r.append('balanced readability')
        scores['lr'] = round(s, 3)
        reasons['lr'] = 'Highly reliable for ' + ', '.join(r) if r else 'Balanced statistical analysis'

        # SVM: complex structure, hedging
        s, r = 0.5, []
        if sl > 20:       s += 0.2; r.append('long complex sentences')
        if cx < 30:       s += 0.15; r.append('complex language')
        if hg > 0.02:     s += 0.15; r.append('hedging language')
        if wc > 40:       s += 0.1; r.append('detailed statement')
        scores['svm'] = round(s, 3)
        reasons['svm'] = 'Detects ' + ', '.join(r) if r else 'Linguistic pattern analysis'

        # DL: medium-length with emotive+structural mix (§3.3.2)
        s, r = 0.5, []
        if 20 < wc < 60:  s += 0.2; r.append('medium-length text')
        if sx > 0.3:      s += 0.15; r.append('moderate sentiment')
        if em > 0.01:     s += 0.15; r.append('emotive language detected')
        if hg > 0.01:     s += 0.1; r.append('hedging cues')
        scores['dl'] = round(s, 3)
        reasons['dl'] = 'Deep encoder strength: ' + ', '.join(r) if r else 'Dual text encoder analysis'

        # RF: diverse features
        s, r = 0.55, []
        if wc > 25:       s += 0.1; r.append('sufficient text length')
        if nd > 0.03:     s += 0.1; r.append('numerical features')
        if 0.01 < em < 0.05: s += 0.1; r.append('moderate emotion')
        s += 0.05
        scores['rf'] = round(s, 3)
        reasons['rf'] = 'Robust with ' + ', '.join(r) if r else 'Robust with diverse text patterns'

        # Ensemble: always safest
        s, r = 0.6, []
        ms = list(scores.values())
        if ms and (max(ms) - min(ms) < 0.2):
            s += 0.2; r.append('models agree')
        if wc > 50:       s += 0.1; r.append('complex statement')
        s += 0.05
        scores['ensemble'] = round(s, 3)
        reasons['ensemble'] = 'Safest: ' + ', '.join(r) if r else 'Combines all model perspectives'

        rec = max(scores, key=scores.get)
        mi = {
            'nb':       {'name': 'Naive Bayes',       'icon': '🔍', 'strength': 'Keyword Spotting'},
            'lr':       {'name': 'Logistic Regression','icon': '⚖️', 'strength': 'Statistical Balance'},
            'svm':      {'name': 'Linear SVC',         'icon': '🔬', 'strength': 'Pattern Recognition'},
            'dl':       {'name': 'Deep Learning',      'icon': '🧠', 'strength': 'Dual Text Encoding (§3.3.2)'},
            'rf':       {'name': 'Random Forest',      'icon': '🌲', 'strength': 'Feature Diversity'},
            'ensemble': {'name': 'Hybrid Ensemble',    'icon': '🏆', 'strength': 'Combined Intelligence'},
        }
        # Sorted scores array for frontend consumption
        scores_array = sorted([{'model': k, 'score': v} for k, v in scores.items()], key=lambda x: x['score'], reverse=True)
        return jsonify({'recommended': rec, 'recommended_model': rec,
                        'reason': reasons.get(rec, ''),
                        'scores': scores_array, 'scores_map': scores,
                        'reasons': reasons, 'model_info': mi})
    except Exception as e:
        logger.error(f"Recommend error: {traceback.format_exc()}")
        return jsonify({'recommended': 'ensemble', 'reason': 'Analysis unavailable.',
                        'scores': {}, 'reasons': {}, 'model_info': {}})


# ─── AUTHENTICATION ROUTES ────────────────────────────────────────────────────
@app.route('/api/auth/signup', methods=['POST'])
@app.route('/auth/signup', methods=['POST'])
def signup():
    data     = request.json or {}
    username = (data.get('username') or '').strip()
    password = (data.get('password') or '').strip()
    email    = (data.get('email') or '').strip()

    # Validate inputs
    err = validate_username(username)
    if err:
        return jsonify({'error': err}), 400
    err = validate_password(password)
    if err:
        return jsonify({'error': err}), 400
    err = validate_email(email)
    if err:
        return jsonify({'error': err}), 400

    try:
        conn = get_conn()
        cursor = conn.cursor()
        hashed_pw = generate_password_hash(password)
        cursor.execute(
            f'INSERT INTO users (username, email, password) VALUES ({PH}, {PH}, {PH})',
            (username, email, hashed_pw)
        )
        conn.commit()
        return jsonify({'message': 'Account created successfully.'}), 201
    except IntegrityError:
        return jsonify({'error': 'Username already exists.'}), 409
    except Exception as e:
        return jsonify({'error': str(e)}), 500
    finally:
        try:
            conn.close()
        except Exception:
            pass


@app.route('/api/auth/login', methods=['POST'])
@app.route('/auth/login', methods=['POST'])
def login():
    ip = request.remote_addr or '0.0.0.0'
    if not _check_rate_limit(ip):
        return jsonify({'error': 'Too many login attempts. Please wait 5 minutes.'}), 429

    data     = request.json or {}
    username = (data.get('username') or '').strip()
    password = (data.get('password') or '').strip()

    if not username or not password:
        return jsonify({'error': 'Username and password are required.'}), 400

    conn = None
    try:
        conn   = get_conn()
        cursor = conn.cursor()
        cursor.execute(f'SELECT password FROM users WHERE username = {PH}', (username,))
        user   = cursor.fetchone()

        if user and check_password_hash(user[0], password):
            token = _issue_token(username)
            return jsonify({
                'message':  'Login successful',
                'username': username,
                'token':    token
            }), 200

        return jsonify({'error': 'Invalid username or password.'}), 401
    except Exception as e:
        return jsonify({'error': str(e)}), 500
    finally:
        try:
            if conn:
                conn.close()
        except Exception:
            pass


@app.route('/api/auth/verify-token', methods=['POST'])
@app.route('/auth/verify-token', methods=['POST'])
def verify_token():
    """Validates a JWT token. Returns username if valid."""
    token = None
    auth = request.headers.get('Authorization', '')
    if auth.startswith('Bearer '):
        token = auth[7:]
    elif request.json:
        token = request.json.get('token')

    if not token:
        return jsonify({'error': 'No token provided.'}), 401

    try:
        username = _decode_token(token)
        return jsonify({'valid': True, 'username': username}), 200
    except jwt.ExpiredSignatureError:
        return jsonify({'valid': False, 'error': 'Token expired. Please log in again.'}), 401
    except jwt.PyJWTError:
        return jsonify({'valid': False, 'error': 'Invalid token.'}), 401


@app.route('/api/auth/change-password', methods=['POST'])
@app.route('/auth/change-password', methods=['POST'])
def change_password():
    data            = request.json or {}
    username        = (data.get('username') or '').strip()
    current_password = (data.get('currentPassword') or '').strip()
    new_password    = (data.get('newPassword') or '').strip()

    if not username or not current_password or not new_password:
        return jsonify({'error': 'All fields required.'}), 400

    err = validate_password(new_password)
    if err:
        return jsonify({'error': err}), 400

    conn = None
    try:
        conn   = get_conn()
        cursor = conn.cursor()
        cursor.execute(f'SELECT password FROM users WHERE username = {PH}', (username,))
        user   = cursor.fetchone()

        if user and check_password_hash(user[0], current_password):
            hashed_pw = generate_password_hash(new_password)
            cursor.execute(f'UPDATE users SET password = {PH} WHERE username = {PH}',
                           (hashed_pw, username))
            conn.commit()
            return jsonify({'message': 'Password changed successfully.'}), 200

        return jsonify({'error': 'Incorrect current password.'}), 401
    except Exception as e:
        return jsonify({'error': str(e)}), 500
    finally:
        try:
            if conn:
                conn.close()
        except Exception:
            pass


# ─── ROOT ROUTE ──────────────────────────────────────────────────────────────
@app.route('/')
def index():
    return jsonify({'status': 'Truth Detector API is running. Frontend is on Vercel.'})


# ─── Fact Check API Proxy ───────────────────────────────────────────────────
@app.route('/api/verify', methods=['POST'])
@app.route('/verify', methods=['POST'])
def verify_claim():
    """
    Proxies the request to Google Fact Check Tools API.
    """
    token = _get_request_token()
    if not token:
        return jsonify({"error": "Unauthorized"}), 401

    data = request.get_json()
    query = data.get('statement')
    if not query:
        return jsonify({"error": "Statement is required"}), 400

    api_key = os.environ.get('GOOGLE_FACT_CHECK_API_KEY')
    if not api_key:
        logger.warning("GOOGLE_FACT_CHECK_API_KEY not set. Using dry-run/mock behavior.")
        # Return a helpful mock response pointing to real documentation if key is missing
        return jsonify({
            "status": "mock",
            "message": "Fact-Check API Key not configured on server.",
            "results": [
                {
                    "claimReview": [
                        {
                            "publisher": {"name": "Veracity System"},
                            "textualRating": "API Key Required",
                            "title": "How to enable live fact-checking"
                        }
                    ],
                    "text": f"Search for: '{query}'"
                }
            ]
        })

    try:
        url = "https://factchecktools.googleapis.com/v1alpha1/claims:search"
        params = {
            "query": query,
            "key": api_key,
            "languageCode": "en"
        }
        resp = requests.get(url, params=params, timeout=10)
        resp.raise_for_status()
        
        results = resp.json()
        return jsonify({
            "status": "success",
            "results": results.get('claims', [])
        })

    except Exception as e:
        logger.error(f"Fact Check API Error: {str(e)}")
        return jsonify({"error": str(e)}), 500


# ─── Startup ─────────────────────────────────────────────────────────────────
init_db()      # Ensure DB exists before accepting requests
load_models()

# ─── Gradio + Flask Integration via FastAPI mount ──────────────────────────────
# The official way to combine Gradio and custom APIs in HF Spaces is to create
# a FastAPI app, mount the custom API, and then mount Gradio on top.
try:
    import gradio as gr
    import spaces
    from fastapi import FastAPI
    from fastapi.middleware.wsgi import WSGIMiddleware

    # ── 1. Gradio UI ─────────────────────────────────────────────────────────
    @spaces.GPU
    def predict_gradio(statement, model_type):
        if not statement or len(statement) < 5:
            return "Please enter a statement with at least 5 characters."
        if not extractor:
            return "Extractor unavailable."
        metadata = {'party': 'independent', 'barely_true_counts': 0, 'false_counts': 0,
                    'half_true_counts': 0, 'mostly_true_counts': 0, 'pants_on_fire_counts': 0}
        raw_feats = extractor.get_combined_features(statement, metadata)
        input_data = {**raw_feats, 'statement': statement}
        input_df = pd.DataFrame([input_data])
        if model_type == 'ensemble' and 'ensemble' not in models and len(models) > 0:
            probs = [m.predict_proba(input_df)[0].tolist()[1] for m in models.values() if hasattr(m, 'predict_proba')]
            real_prob = float(np.mean(probs)) if probs else 0.5
        else:
            m = models.get(model_type) or (list(models.values())[0] if models else None)
            if not m:
                return "Model unavailable."
            try:
                real_prob = m.predict_proba(input_df)[0].tolist()[1]
            except Exception:
                real_prob = 0.5
        t_data = thresholds.get(model_type, {})
        threshold = t_data.get('threshold', 0.5) if t_data else 0.5
        pred = 1 if real_prob >= threshold else 0
        label = 'Real' if pred == 1 else 'Fake'
        conf = real_prob * 100 if pred == 1 else (1 - real_prob) * 100
        return f"Prediction: {label} ({conf:.2f}% confidence)\nReal Prob: {real_prob*100:.2f}%\nFake Prob: {(1-real_prob)*100:.2f}%"

    demo = gr.Interface(
        fn=predict_gradio,
        inputs=[
            gr.Textbox(lines=4, placeholder="Enter statement to verify...", label="News Statement"),
            gr.Dropdown(choices=['ensemble', 'lr', 'nb', 'svm', 'rf', 'dl'], value='ensemble', label="Model")
        ],
        outputs="text",
        title="Truth Detector API & Interactive Demo",
        description="Backend API for Fake News Detection. REST API available at /api/*"
    )

    demo.app.mount("/flask", WSGIMiddleware(app))

    _GRADIO_AVAILABLE = True

except ImportError as _ie:
    _GRADIO_AVAILABLE = False
    logger.info(f"Gradio not available ({_ie}) — running in Flask-only (Docker) mode.")

if __name__ == '__main__':
    port = int(os.environ.get('PORT', 7860))
    if _GRADIO_AVAILABLE:
        demo.launch(server_name="0.0.0.0", server_port=port)
    else:
        app.run(host='0.0.0.0', port=port, debug=False)