app.py

import re
import gradio as gr
from transformers import T5ForConditionalGeneration, T5Tokenizer
from difflib import SequenceMatcher

# ----------------- Load Model -----------------
model_id = "jang17/AGLI_model"  # Your model repo
tokenizer = T5Tokenizer.from_pretrained(model_id)
model = T5ForConditionalGeneration.from_pretrained(model_id)

# ----------------- Rule-Based Pre-Processor -----------------
def apply_rules(text: str) -> str:
    if not text:
        return text

    # 1. Strip & collapse internal whitespace
    text = text.strip()
    text = re.sub(r' {2,}', ' ', text)

    # 2. Remove space before punctuation
    text = re.sub(r'\s+([.,!?;:])', r'\1', text)

    # 3. Ensure one space after punctuation (but not at end-of-string)
    text = re.sub(r'([.,!?;:])(?=[^\s])', r'\1 ', text)

    # 4. Capitalize first character
    text = text[0].upper() + text[1:] if len(text) > 1 else text.upper()

    # 5. Capitalize letter after terminal punctuation
    text = re.sub(
        r'([.!?])\s+([a-z])',
        lambda m: m.group(1) + ' ' + m.group(2).upper(),
        text
    )

    # 6. Capitalize standalone 'i'
    text = re.sub(r'\bi\b', 'I', text)

    # 7. Add terminal period if missing
    if text and text[-1] not in '.!?':
        text += '.'

    return text

# ----------------- Grammar Correction Function -----------------
def correct_grammar(text):
    original_text = text.strip()
    if not original_text:
        return "", "", []

    # Apply rules before T5
    rule_corrected = apply_rules(original_text)

    inputs = tokenizer(
        "grammar: " + rule_corrected,
        return_tensors="pt",
        max_length=256,
        truncation=True
    )
    outputs = model.generate(
        **inputs,
        max_length=256,
        num_beams=6,
        early_stopping=False
    )
    corrected = tokenizer.decode(outputs[0], skip_special_tokens=True)

    # Apply rules after T5
    corrected = apply_rules(corrected)

    original_words = original_text.split()
    corrected_words = corrected.split()

    matcher = SequenceMatcher(None, original_words, corrected_words)
    opcodes = list(matcher.get_opcodes())

    # Build initial groups (merge gaps ≤1 token)
    groups = []
    current = None
    for idx, (tag, i1, i2, j1, j2) in enumerate(opcodes):
        if tag != "equal":
            if current is None:
                current = [i1, i2, j1, j2]
            else:
                current[1] = i2
                current[3] = j2
        else:
            gap = i2 - i1
            next_is_change = idx + 1 < len(opcodes) and opcodes[idx + 1][0] != "equal"
            if current is not None and gap <= 1 and next_is_change:
                current[1] = i2
                current[3] = j2
            else:
                if current is not None:
                    groups.append(current)
                    current = None
    if current is not None:
        groups.append(current)

    # Merge groups that share words (word-order / move detection)
    merged = True
    while merged:
        merged = False
        skip = set()
        for i in range(len(groups)):
            if i in skip:
                continue
            g1 = groups[i]
            g1_orig = set(original_words[g1[0]:g1[1]])
            g1_corr = set(corrected_words[g1[2]:g1[3]])
            for j in range(i + 1, len(groups)):
                if j in skip:
                    continue
                g2 = groups[j]
                g2_orig = set(original_words[g2[0]:g2[1]])
                g2_corr = set(corrected_words[g2[2]:g2[3]])
                if g1_orig & g2_corr or g1_corr & g2_orig:
                    groups[i] = [
                        min(g1[0], g2[0]), max(g1[1], g2[1]),
                        min(g1[2], g2[2]), max(g1[3], g2[3])
                    ]
                    g1 = groups[i]
                    g1_orig = set(original_words[g1[0]:g1[1]])
                    g1_corr = set(corrected_words[g1[2]:g1[3]])
                    skip.add(j)
                    merged = True
        groups = [g for i, g in enumerate(groups) if i not in skip]

    errors = []
    for i1, i2, j1, j2 in groups:
        orig_part = " ".join(original_words[i1:i2])
        corr_part = " ".join(corrected_words[j1:j2])
        if orig_part or corr_part:
            errors.append({
                "original": orig_part,
                "correction": corr_part
            })

    return original_text, corrected, errors

# ----------------- Gradio Interface -----------------
demo = gr.Interface(
    fn=correct_grammar,
    inputs=gr.Textbox(
        label="Enter text to correct",
        lines=5,
        placeholder="Type or paste your sentence/paragraph here..."
    ),
    outputs=[
        gr.Textbox(label="Original Text"),
        gr.Textbox(label="Corrected Text"),
        gr.JSON(label="Errors")
    ],
    title="AGLI Grammar Correction",
    description="Fine-tuned T5 model for grammar correction. Paste text and get corrections + detected errors.",
    examples=[
        ["He go to school yesterday but forget his book."],
        ["The childrens plays happy in the parks."],
        ["She don't likes apple but she love oranges."],
        ["I have cat."]
    ],
    flagging_mode="never"
)

# ----------------- Launch -----------------
if __name__ == "__main__":
    demo.launch()