# ============================================================
#  TruthLens — AI Fake News Detector (High Accuracy & Simplified)
#  HuggingFace Spaces compatible (Gradio native components)
#  Stack: HuggingFace Transformers + DuckDuckGo Search + Gradio
# ============================================================

import re
import json
from urllib.parse import urlparse
import numpy as np
import torch
import gradio as gr
from PIL import Image, ImageEnhance, ImageOps
import pytesseract
from transformers import (
    pipeline,
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
)
from duckduckgo_search import DDGS

# ─────────────────────────────────────────
# 1.  LOAD MODELS
# ─────────────────────────────────────────
print("Loading summariser (facebook/bart-large-cnn)...")
_SUM_NAME      = "facebook/bart-large-cnn"
_sum_tokenizer = AutoTokenizer.from_pretrained(_SUM_NAME)
_sum_model     = AutoModelForSeq2SeqLM.from_pretrained(_SUM_NAME)
_sum_model.eval()

print("Loading fake-news classifier...")
classifier = pipeline(
    "text-classification",
    model="hamzab/roberta-fake-news-classification",
    truncation=True,
)

print("Loading zero-shot classifier (facebook/bart-large-mnli)...")
zs = pipeline(
    "zero-shot-classification",
    model="facebook/bart-large-mnli",
)

print("All models ready.")

# ─────────────────────────────────────────
# 2.  DOMAIN CREDIBILITY REGISTRY
# ─────────────────────────────────────────
TRUSTED = {
    "reuters.com": 10, "apnews.com": 10, "bbc.com": 9, "bbc.co.uk": 9,
    "snopes.com": 10, "factcheck.org": 10, "politifact.com": 10,
    "fullfact.org": 9, "afp.com": 9,
    "nytimes.com": 8, "washingtonpost.com": 8, "theguardian.com": 8,
    "npr.org": 8, "pbs.org": 8, "theatlantic.com": 7,
    "economist.com": 8, "ft.com": 8, "bloomberg.com": 7,
    "time.com": 7, "forbes.com": 6, "cnn.com": 6,
    "nbcnews.com": 6, "abcnews.go.com": 6, "cbsnews.com": 6,
    "ndtv.com": 6, "thehindu.com": 7, "hindustantimes.com": 6,
    "indiatoday.in": 6, "wikipedia.org": 5,
}

SATIRE       = ["theonion.com","babylonbee.com","clickhole.com",
                "worldnewsdailyreport.com","waterfordwhispersnews.com"]
DISREPUTABLE = ["infowars.com","naturalnews.com","beforeitsnews.com"]

def domain_score(url):
    try:
        d = urlparse(url).netloc.lower().replace("www.", "")
    except Exception:
        return 0, "Unknown"

    for bad in SATIRE:
        if bad in d: return -8, "Satire"
    for bad in DISREPUTABLE:
        if bad in d: return -6, "Disreputable"
    for good, s in TRUSTED.items():
        if d.endswith(good) or good in d:
            if s >= 9: return s, "Premier"
            if s >= 7: return s, "Credible"
            return s, "Mainstream"
    return 2, "Unknown"

# ─────────────────────────────────────────
# 3.  TEXT UTILITIES
# ─────────────────────────────────────────
STOP = {
    "the","a","an","is","was","are","were","in","on","at","to","of",
    "and","or","for","with","that","this","it","as","by","from",
    "has","have","had","be","been","not","but","so","do","did","will",
}

def clean_text(text):
    text = re.sub(r'http\S+|www\.\S+|\S+@\S+', '', text)
    text = re.sub(r'[^a-zA-Z0-9\s.,!?\'\"%-]', ' ', text)
    return re.sub(r'\s+', ' ', text).strip()

def build_query(text):
    # ACCURACY FIX 1: Removed exact quotes so search works better
    words = re.findall(r'\b[a-zA-Z0-9]{3,}\b', text)
    meaningful = [w for w in words if w.lower() not in STOP]
    core = " ".join(meaningful[:12])
    return f"{core} news"

def ocr_image(image):
    if image is None: return ""
    if not isinstance(image, Image.Image):
        image = Image.fromarray(np.uint8(image))
    
    image = ImageOps.grayscale(image)
    image = ImageEnhance.Contrast(image).enhance(2.5)
    image = image.point(lambda x: 0 if x < 140 else 255)
    return clean_text(pytesseract.image_to_string(image, config='--oem 3 --psm 6'))

# ─────────────────────────────────────────
# 4.  SUMMARISE
# ─────────────────────────────────────────
def summarise(text):
    words = text.split()
    if len(words) < 30: return text
    chunk = " ".join(words[:800])
    
    try:
        inputs = _sum_tokenizer(chunk, return_tensors="pt", max_length=1024, truncation=True)
        with torch.no_grad():
            ids = _sum_model.generate(
                inputs["input_ids"],
                max_new_tokens=80, min_length=25, num_beams=4, early_stopping=True,
            )
        return _sum_tokenizer.decode(ids[0], skip_special_tokens=True)
    except Exception:
        return " ".join(words[:40])

# ─────────────────────────────────────────
# 5.  WEB SEARCH
# ─────────────────────────────────────────
def web_search(summary_text):
    query = build_query(summary_text)
    try:
        raw = list(DDGS().text(query, max_results=7))
    except Exception as e:
        print(f"DuckDuckGo API Error: {e}") 
        return [], f"Search error: {e}", 0

    sources, total = [], 0
    for r in raw:
        link     = r.get("href", "")
        sc, tier = domain_score(link)
        total   += max(0, sc)
        sources.append({
            "title": r.get("title", "—"),
            "body":  r.get("body", "")[:240],
            "link":  link,
            "score": sc,
            "tier":  tier,
        })
    sources.sort(key=lambda x: x["score"], reverse=True)
    return sources, query, total

# ─────────────────────────────────────────
# 6.  VERDICT ENGINE
# ─────────────────────────────────────────
def verdict_engine(text, web_total, sources):
    score, signals = 50, []

    # --- 1. AI Classifier ---
    try:
        r    = classifier(text[:512])[0]
        lbl  = r["label"].upper() 
        conf = round(r["score"] * 100, 1)

        if lbl in ["REAL", "TRUE"]: 
            pts = round(conf / 100 * 40)
            score += pts
            signals.append(("🤖 AI Classifier", f"REAL ({conf}%)", f"+{pts}", "pos"))
        else:
            pts = round(conf / 100 * 40)
            score -= pts
            signals.append(("🤖 AI Classifier", f"FAKE ({conf}%)", f"-{pts}", "neg"))
    except Exception:
        signals.append(("🤖 AI Classifier", "Skipped", "±0", "neu"))

    # --- 2. Zero-Shot NLI ---
    try:
        zr   = zs(text[:512], ["credible news", "misinformation", "satire"])
        top  = zr["labels"][0]
        zcon = round(zr["scores"][0] * 100, 1)

        if top == "credible news":
            score += 15
            signals.append(("🧠 Zero-Shot NLI", f"Credible ({zcon}%)", "+15", "pos"))
        elif top == "misinformation":
            score -= 15
            signals.append(("🧠 Zero-Shot NLI", f"Misinformation ({zcon}%)", "-15", "neg"))
        else:
            score -= 8
            signals.append(("🧠 Zero-Shot NLI", f"Satire ({zcon}%)", "-8", "neg"))
    except Exception:
        signals.append(("🧠 Zero-Shot NLI", "Skipped", "±0", "neu"))

    # --- 3. Web Coverage ---
    if not sources:
        # ACCURACY FIX 2: Lowered penalty from -15 to -5 when API fails
        score -= 5
        signals.append(("🌐 Web Coverage", "No sources found", "-5", "neg"))
    elif web_total >= 20:
        pts    = min(35, web_total * 2)
        score += pts
        signals.append(("🌐 Web Coverage", f"Strong ({web_total} pts)", f"+{pts}", "pos"))
    elif web_total >= 8:
        score += 10
        signals.append(("🌐 Web Coverage", f"Moderate ({web_total} pts)", "+10", "pos"))
    else:
        score -= 10
        signals.append(("🌐 Web Coverage", f"Weak ({web_total} pts)", "-10", "neg"))

    score = max(0, min(100, score))

    if score >= 70:   tag, cls = "✅ LIKELY REAL", "real"
    elif score >= 45: tag, cls = "⚠️ UNCERTAIN",   "uncertain"
    else:             tag, cls = "🚨 LIKELY FAKE", "fake"

    return score, tag, cls, signals

# ─────────────────────────────────────────
# 7.  FORMAT OUTPUT AS HTML (SIMPLIFIED)
# ─────────────────────────────────────────
def build_html(score, tag, cls, signals, summary, query, sources):
    cls_colors = {
        "real":      ("#2ecc71", "rgba(46,204,113,0.08)"),
        "fake":      ("#e05252", "rgba(224,82,82,0.08)"),
        "uncertain": ("#d4a843", "rgba(212,168,67,0.08)"),
    }
    color, bg = cls_colors.get(cls, cls_colors["uncertain"])

    desc = {
        "real": "Content appears credible based on AI analysis.",
        "fake": "Significant misinformation indicators detected.",
        "uncertain": "Evidence is mixed — manual verification recommended."
    }.get(cls, "")

    html = f"""
    <div style="font-family:'Lora',Georgia,serif;color:#e8e2d9;background:#0d0d0f;
                border-radius:12px;overflow:hidden;border:1px solid #1e1e26;
                padding:40px 30px; text-align:center;">
        
        <div style="background:{bg}; border:2px solid {color}; border-radius:10px;
                    padding:30px; display:inline-block; min-width:300px;">
            
            <div style="font-family:'Syne',sans-serif;font-weight:800;font-size:2.8rem;
                        letter-spacing:-.03em;color:{color}; line-height:1;">
                {tag}
            </div>
            
            <div style="border:none;border-top:1px solid #1e1e26;width:80px;margin:20px auto;"></div>

            <div style="font-family:'Lora',serif;font-size:16px;color:#a09898; 
                        line-height:1.6; max-width:400px; margin:0 auto;">
                {desc}
            </div>
        </div>

    </div>
    """
    return html

# ─────────────────────────────────────────
# 8.  MAIN PIPELINE
# ─────────────────────────────────────────
def analyse(text, image):
    ocr_text = ""
    if image is not None:
        ocr_text = ocr_image(image)
        if len(ocr_text.split()) < 3:
            ocr_text = ""
    
    provided_text = clean_text(text or "")

    raw = ""
    if ocr_text and provided_text:
        raw = f"{ocr_text}. {provided_text}"
    elif ocr_text:
        raw = ocr_text
    else:
        raw = provided_text

    if len(raw.split()) < 3:
        return build_error_html("Input too short. Please provide a headline or sentence.")

    summary            = summarise(raw)
    sources, query, wt = web_search(summary)
    score, tag, cls, signals = verdict_engine(raw, wt, sources)

    return build_html(score, tag, cls, signals, summary, query, sources)

def build_error_html(msg):
    return f"""<div style="background:rgba(224,82,82,0.07);border:1px solid rgba(224,82,82,0.25);
            border-radius:8px;padding:20px 22px;color:#e07878;
            font-family:'DM Mono',monospace;font-size:13px; text-align:center;">
  ⚠️ &nbsp;{msg}
</div>"""

# ─────────────────────────────────────────
# 9.  GRADIO UI
# ─────────────────────────────────────────
GOOGLE_FONTS = """@import url('https://fonts.googleapis.com/css2?family=Syne:wght@400;600;700;800&family=DM+Mono:wght@400;500&family=Lora:ital,wght@0,400;1,400&display=swap');"""

CUSTOM_CSS = GOOGLE_FONTS + """
/* ── RESET & BASE ── */
*, *::before, *::after { box-sizing: border-box; }
body, .gradio-container, #root {
    background: #0a0a0e !important;
    font-family: 'Lora', Georgia, serif !important;
    color: #e8e2d9 !important;
    min-height: 100vh;
}
.gradio-container {
    max-width: 1100px !important;
    margin: 0 auto !important;
    padding: 0 !important;
}

/* ── HEADER ── */
#tl-header-html {
    text-align: center;
    padding: 52px 24px 36px;
    border-bottom: 1px solid #1e1e26;
    background: linear-gradient(180deg, #050508 0%, #0a0a0e 100%);
    margin-bottom: 0;
}

/* ── PANELS / BOXES ── */
.tl-input-col, .tl-result-col {
    background: #111116;
    border: 1px solid #1e1e26;
    border-radius: 10px;
    padding: 26px 28px;
}

/* ── GRADIO OVERRIDES ── */
.gr-form label, label.svelte-1b6s6s, .block > label {
    font-family: 'DM Mono', monospace !important;
    font-size: 9px !important;
    letter-spacing: .3em !important;
    text-transform: uppercase !important;
    color: #4a4a56 !important;
    margin-bottom: 10px !important;
}

textarea, .gr-textbox textarea, input[type="text"] {
    background: #141418 !important;
    border: 1px solid #252530 !important;
    border-radius: 7px !important;
    color: #e8e2d9 !important;
    font-family: 'Lora', serif !important;
    font-size: 14px !important;
    line-height: 1.75 !important;
    padding: 14px 16px !important;
    transition: border-color .2s, box-shadow .2s !important;
    resize: vertical !important;
}
textarea:focus, input[type="text"]:focus {
    border-color: #3a3a50 !important;
    box-shadow: 0 0 0 3px rgba(224,82,82,0.07) !important;
    outline: none !important;
}
textarea::placeholder { color: #3a3a44 !important; }

.gr-image, .image-container, [data-testid="image"] {
    background: #141418 !important;
    border: 1.5px dashed #252530 !important;
    border-radius: 7px !important;
    transition: border-color .2s !important;
}
.gr-image:hover { border-color: #e05252 !important; }

#tl-analyse-btn, button#tl-analyse-btn {
    background: #e05252 !important;
    color: white !important;
    font-family: 'Syne', sans-serif !important;
    font-weight: 700 !important;
    font-size: 13px !important;
    letter-spacing: .18em !important;
    text-transform: uppercase !important;
    border: none !important;
    border-radius: 7px !important;
    padding: 15px 24px !important;
    cursor: pointer !important;
    transition: background .2s, box-shadow .2s !important;
    width: 100% !important;
    margin-top: 6px !important;
}
#tl-analyse-btn:hover {
    background: #c94444 !important;
    box-shadow: 0 0 28px rgba(224,82,82,0.35) !important;
}

#tl-clear-btn {
    background: transparent !important;
    color: #4a4a56 !important;
    font-family: 'DM Mono', monospace !important;
    font-size: 11px !important;
    letter-spacing: .12em !important;
    text-transform: uppercase !important;
    border: 1px solid #252530 !important;
    border-radius: 7px !important;
    padding: 10px 24px !important;
    cursor: pointer !important;
    transition: all .2s !important;
    width: 100% !important;
}
#tl-clear-btn:hover {
    border-color: #3a3a50 !important;
    color: #e8e2d9 !important;
}

.gr-html, [data-testid="html"] {
    background: transparent !important;
    border: none !important;
    padding: 0 !important;
}

.gr-accordion {
    background: #111116 !important;
    border: 1px solid #1e1e26 !important;
    border-radius: 8px !important;
}

footer { display: none !important; }
.built-with { display: none !important; }
#share-btn-container { display: none !important; }

.gap-4 { gap: 20px !important; }

::-webkit-scrollbar { width: 6px; }
::-webkit-scrollbar-track { background: #0a0a0e; }
::-webkit-scrollbar-thumb { background: #252530; border-radius: 3px; }
"""

HEADER_HTML = """
<link rel="preconnect" href="https://fonts.googleapis.com">
<link href="https://fonts.googleapis.com/css2?family=Syne:wght@400;600;700;800&family=DM+Mono:wght@400;500&family=Lora:ital,wght@0,400;1,400&display=swap" rel="stylesheet">
<div id="tl-header-html" style="
    text-align:center;
    padding:52px 24px 36px;
    border-bottom:1px solid #1e1e26;
    background:linear-gradient(180deg,#050508 0%,#0a0a0e 100%);
    position:relative;overflow:hidden;">
  <div style="position:absolute;top:-80px;left:50%;transform:translateX(-50%);
              width:500px;height:220px;
              background:radial-gradient(ellipse,rgba(224,82,82,0.06) 0%,transparent 70%);
              pointer-events:none;"></div>
  
  <div style="font-family:'DM Mono',monospace;font-size:10px;letter-spacing:.5em;
              color:#3a3a44;text-transform:uppercase;margin-bottom:14px;">
    — Powered by AI —
  </div>
  <div style="font-family:'Syne',sans-serif;font-size:clamp(2.4rem,5vw,4rem);
              font-weight:800;letter-spacing:-.03em;line-height:1;color:#e8e2d9;">
    Truth<span style="color:#e05252;">Lens</span>
  </div>
  <div style="border:none;border-top:1px solid #1e1e26;width:56px;margin:16px auto 12px;"></div>
  <div style="font-style:italic;color:#4a4a56;font-size:15px;font-family:'Lora',serif;">
    AI-powered misinformation detector
  </div>
</div>
"""

EXAMPLES = [
    ["Scientists at MIT have discovered that drinking coffee daily can reverse the aging process in humans.", None],
    ["NASA confirms water ice found on the surface of Mars in a new landmark discovery published in Science journal.", None],
    ["The government is secretly adding mind-control chemicals to public drinking water to suppress dissent among citizens.", None],
]

PLACEHOLDER = (
    "Paste a news headline or any claim you want to verify...\n\n"
    "Tip: You can also upload a screenshot of a news article."
)

with gr.Blocks(
    title="TruthLens — AI Fake News Detector",
    css=CUSTOM_CSS,
    theme=gr.themes.Base(
        primary_hue=gr.themes.colors.red,
        neutral_hue=gr.themes.colors.slate,
        font=[gr.themes.GoogleFont("Lora"), "Georgia", "serif"],
    ),
) as app:

    gr.HTML(HEADER_HTML)

    with gr.Row(equal_height=False):
        
        with gr.Column(scale=4, min_width=320):
            gr.HTML("""
            <div style="font-family:'DM Mono',monospace;font-size:9px;letter-spacing:.3em;
                        text-transform:uppercase;color:#4a4a56;margin:20px 0 10px;
                        display:flex;align-items:center;gap:10px;">
              News Text or Headline
              <span style="flex:1;height:1px;background:#1e1e26;display:block;"></span>
            </div>""")
            
            text_input = gr.Textbox(
                lines=7,
                max_lines=16,
                placeholder=PLACEHOLDER,
                show_label=False,
                container=False,
            )

            gr.HTML("""
            <div style="font-family:'DM Mono',monospace;font-size:9px;letter-spacing:.3em;
                        text-transform:uppercase;color:#4a4a56;margin:18px 0 10px;
                        display:flex;align-items:center;gap:10px;">
              Or Upload Screenshot (OCR)
              <span style="flex:1;height:1px;background:#1e1e26;display:block;"></span>
            </div>""")

            image_input = gr.Image(
                type="pil",
                label="Upload screenshot",
                show_label=False,
                container=False,
                height=180,
            )

            gr.HTML('<div style="height:12px;"></div>')

            analyse_btn = gr.Button(
                "→  Analyse Claim",
                variant="primary",
                elem_id="tl-analyse-btn",
            )
            
            clear_btn = gr.Button(
                "✕  Clear",
                variant="secondary",
                elem_id="tl-clear-btn",
            )
            
            gr.HTML("""
            <div style="margin-top:20px;padding:10px 14px;background:#0e0e12;
                        border:1px solid #1e1e26;border-radius:7px;
                        font-family:'DM Mono',monospace;font-size:10px;color:#3a3a44;
                        text-align:center;letter-spacing:.06em;">
              ⚠️  AI verdict is advisory only
            </div>""")

        with gr.Column(scale=6, min_width=400):
            gr.HTML("""
            <div style="font-family:'DM Mono',monospace;font-size:9px;letter-spacing:.3em;
                        text-transform:uppercase;color:#4a4a56;margin:20px 0 10px;
                        display:flex;align-items:center;gap:10px;">
              Analysis Result
              <span style="flex:1;height:1px;background:#1e1e26;display:block;"></span>
            </div>""")
            
            result_output = gr.HTML(
                value="""
                <div style="text-align:center;padding:72px 24px;color:#2a2a34;
                            border:1px dashed #1e1e26;border-radius:10px;background:#0e0e12;">
                  <div style="font-size:3rem;margin-bottom:16px;opacity:.3;">🔍</div>
                  <div style="font-family:'DM Mono',monospace;font-size:12px;letter-spacing:.1em;
                              text-transform:uppercase;">
                    Paste a claim and press Analyse
                  </div>
                </div>""",
                show_label=False,
                container=False,
            )

    gr.HTML("""
    <div style="font-family:'DM Mono',monospace;font-size:9px;letter-spacing:.3em;
                text-transform:uppercase;color:#4a4a56;margin:28px 0 12px;
                display:flex;align-items:center;gap:10px;">
      Try These Examples
      <span style="flex:1;height:1px;background:#1e1e26;display:block;"></span>
    </div>""")

    gr.Examples(
        examples=EXAMPLES,
        inputs=[text_input, image_input],
        outputs=result_output,
        fn=analyse,
        cache_examples=False,
        label=None,
    )

    analyse_btn.click(
        fn=analyse,
        inputs=[text_input, image_input],
        outputs=result_output,
        show_progress="minimal",
    )

    def clear_all():
        empty = """
        <div style="text-align:center;padding:72px 24px;color:#2a2a34;
                    border:1px dashed #1e1e26;border-radius:10px;background:#0e0e12;">
          <div style="font-size:3rem;margin-bottom:16px;opacity:.3;">🔍</div>
          <div style="font-family:'DM Mono',monospace;font-size:12px;letter-spacing:.1em;
                      text-transform:uppercase;">
            Paste a claim and press Analyse
          </div>
        </div>"""
        return "", None, empty

    clear_btn.click(
        fn=clear_all,
        inputs=[],
        outputs=[text_input, image_input, result_output],
    )

    app.load(
        fn=None,
        js="""
        () => {
            document.addEventListener('keydown', function(e) {
                if ((e.ctrlKey || e.metaKey) && e.key === 'Enter') {
                    const btn = document.getElementById('tl-analyse-btn');
                    if (btn) btn.click();
                }
            });
        }
        """
    )

if __name__ == "__main__":
    app.launch()