app.py

import gradio as gr
import requests
from PIL import Image
import pytesseract
import difflib
from io import BytesIO
from transformers import pipeline
import trafilatura
from nltk.tokenize import sent_tokenize
import nltk

nltk.download("punkt")

# === Load AI model ===
reviewer = pipeline("text-generation", model="HuggingFaceH4/zephyr-7b-beta", max_new_tokens=200)
device = "cpu"
print(f"Device set to use {device}")

# === Utility: Highlight diffs ===
def highlight_diff(original, suggestion):
    diff = difflib.ndiff(original.split(), suggestion.split())
    result = ""
    for word in diff:
        if word.startswith("- "):
            result += f"<span style='color:red;text-decoration:line-through'>{word[2:]}</span> "
        elif word.startswith("+ "):
            result += f"<span style='color:green;font-weight:bold'>{word[2:]}</span> "
        elif word.startswith("  "):
            result += word[2:] + " "
    return result.strip()

# === Extract blog content from URL ===
def extract_text_from_url(url):
    try:
        headers = {"User-Agent": "Mozilla/5.0"}
        response = requests.get(url, headers=headers, timeout=10)
        if response.status_code == 200:
            return trafilatura.extract(response.text)
        else:
            return f"❌ Blog Error: HTTP {response.status_code} on URL {url}"
    except Exception as e:
        return f"❌ Blog Error: {e}"

# === Extract text from image URL (OCR) ===
def extract_text_from_image(image_url):
    try:
        img_data = requests.get(image_url).content
        image = Image.open(BytesIO(img_data)).convert("L")
        text = pytesseract.image_to_string(image)
        return text if text.strip() else "❌ OCR Error: No readable text found."
    except Exception as e:
        return f"❌ OCR Error: {e}"

# === Suggestion generator ===
def generate_suggestions(text):
    sentences = sent_tokenize(text)
    suggestions = []
    for sent in sentences:
        prompt = f"Improve the tone, grammar, clarity and flag any sensitive content:\n\n{sent}"
        output = reviewer(prompt, max_new_tokens=200)[0]["generated_text"]
        cleaned = output.replace(prompt, "").strip()
        suggestions.append(cleaned if cleaned else sent)
    return sentences, suggestions

# === Final approval handler ===
def collect_decisions(originals, suggestions, *choices):
    results = []
    for orig, sugg, choice in zip(originals, suggestions, choices):
        results.append(sugg if choice == "Accept" else orig)
    return "\n".join(results)

# === Gradio UI ===
with gr.Blocks() as demo:
    gr.Markdown("# ✨ Blog Reviewer AI")
    gr.Markdown("Detect tone issues, errors, and sensitive content — and clean them interactively!")

    with gr.Tab("🔗 From Blog URL"):
        blog_url = gr.Textbox(label="Enter blog URL")
        fetch_btn = gr.Button("Fetch & Review")

    with gr.Tab("🖼️ From Image URL (OCR)"):
        image_url = gr.Textbox(label="Enter Image URL")
        image_btn = gr.Button("Extract & Review")

    with gr.Tab("📝 Paste Text"):
        pasted_text = gr.Textbox(label="Paste blog content here", lines=10)
        paste_btn = gr.Button("Review Text")

    output_section = gr.Column(visible=False)
    originals = gr.State([])
    suggestions = gr.State([])
    decision_radios = []

    view_mode = gr.Radio(["Original", "Suggestion", "Side-by-Side"], value="Side-by-Side", label="Choose View")
    final_output = gr.Textbox(label="✅ Final Output", lines=12)
    finalize_btn = gr.Button("Generate Clean Version")

    sentence_blocks = []

    # === Show suggestions UI ===
    def show_review(text):
        origs, suggs = generate_suggestions(text)
        originals.value = origs
        suggestions.value = suggs
        return origs, suggs, True

    # === Populate sentence review rows dynamically ===
    def populate_review_ui(origs, suggs):
        global decision_radios, sentence_blocks
        decision_radios = []
        sentence_blocks = []

        ui_blocks = []
        for i, (orig, sugg) in enumerate(zip(origs, suggs)):
            orig_md = gr.Markdown(f"<b>{orig}</b>", visible=False)
            sugg_md = gr.Markdown(f"<b>{sugg}</b>", visible=False)
            diff_md = gr.Markdown(highlight_diff(orig, sugg), visible=True)

            radio = gr.Radio(["Accept", "Reject"], value="Accept", label=f"Suggestion {i+1}")
            decision_radios.append(radio)
            sentence_blocks.append((orig_md, sugg_md, diff_md))

            ui_blocks.extend([orig_md, sugg_md, diff_md, radio])
        return ui_blocks

    # === Toggle view mode ===
    def toggle_view(view):
        updates = []
        for orig_md, sugg_md, diff_md in sentence_blocks:
            if view == "Original":
                updates.extend([gr.update(visible=True), gr.update(visible=False), gr.update(visible=False)])
            elif view == "Suggestion":
                updates.extend([gr.update(visible=False), gr.update(visible=True), gr.update(visible=False)])
            else:  # Side-by-side
                updates.extend([gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)])
        return updates

    # === Final output handler ===
    def finalize_output(origs, suggs, *choices):
        return collect_decisions(origs, suggs, *choices)

    # Button click handlers
    fetch_btn.click(fn=extract_text_from_url, inputs=blog_url, outputs=pasted_text)
    image_btn.click(fn=extract_text_from_image, inputs=image_url, outputs=pasted_text)

    paste_btn.click(fn=show_review, inputs=pasted_text, outputs=[originals, suggestions, output_section])

    # Dynamic render trigger
    originals.change(fn=populate_review_ui, inputs=[originals, suggestions], outputs=[])

    view_mode.change(fn=toggle_view, inputs=view_mode,
                     outputs=[item for block in sentence_blocks for item in block])

    finalize_btn.click(fn=finalize_output, inputs=[originals, suggestions] + decision_radios, outputs=final_output)

demo.launch()