app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import json
import re
import os
from typing import Dict, Any

# System prompt (must match training)
SYSTEM_PROMPT = """Determine if the message is a restaurant reservation request.  
If yes, extract the following three fields as strings:
- "num_people": number of people (as a string, e.g., "4"). If not mentioned, use an empty string ("").
- "reservation_date": the exact date/time phrase from the message (as a string, do not convert or interpret; e.g., keep "this Saturday at 7 PM" as is). If not mentioned, use an empty string ("").
- "phone_num": the phone number (as a string, digits only, remove any hyphens or formatting; e.g., "0912345678"). If not mentioned, use an empty string ("").

If the message is NOT a reservation request, return:
```json
{
  "num_people": "",
  "reservation_date": "",
  "phone_num": ""
}
```

Output must be valid JSON only, with exactly these three fields and no additional text, fields, or explanations.
"""

# Global variables for model caching
model = None
tokenizer = None

def load_model():
    """Load the model and tokenizer with caching"""
    global model, tokenizer
    if model is not None and tokenizer is not None:
        return model, tokenizer
    
    try:
        print("Loading model...")
        model_name = "Luigi/gemma-3-270m-it-dinercall-ner"
        
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForCausalLM.from_pretrained(
            model_name,
            torch_dtype=torch.float32,
            device_map="auto",
            trust_remote_code=True
        )
        
        # Set padding token if not set
        if tokenizer.pad_token is None:
            tokenizer.pad_token = tokenizer.eos_token
            
        print("Model loaded successfully!")
        return model, tokenizer
    except Exception as e:
        print(f"Error loading model: {e}")
        return None, None

def validate_json(output: str) -> tuple:
    """Validate and extract JSON from model output - supports both plain JSON and code block formats"""
    try:
        # First, try to extract JSON from code blocks (new model version)
        json_match = re.search(r'```(?:json)?\s*(\{[\s\S]*?\})\s*```', output)
        if json_match:
            json_str = json_match.group(1)
        else:
            # If no code block, look for JSON directly (old model version)
            json_match = re.search(r'\{[\s\S]*\}', output)
            if not json_match:
                return False, None, "No JSON found / 未找到JSON"
            json_str = json_match.group(0)
        
        # Fix common JSON issues for both formats
        # 1. Add quotes around phone numbers (they often start with 0)
        json_str = re.sub(r'("phone_num":\s*)(\d[-\d]*)', r'\1"\2"', json_str)
        
        # 2. Add quotes around num_people if it's a number
        json_str = re.sub(r'("num_people":\s*)(\d+)', r'\1"\2"', json_str)
        
        # 3. Fix trailing commas
        json_str = re.sub(r',\s*\}', '}', json_str)
        
        parsed = json.loads(json_str)
        return True, parsed, "Valid JSON / 有效的JSON"
    except json.JSONDecodeError:
        return False, None, "Invalid JSON format / 無效的JSON格式"
    except Exception:
        return False, None, "Error parsing JSON / 解析JSON時出錯"

def extract_reservation_info(text: str):
    """Extract reservation information from text"""
    # Load model if not already loaded
    model, tokenizer = load_model()
    
    if model is None or tokenizer is None:
        return {"error": "Model not loaded, please refresh the page / 模型未加載成功，請刷新頁面重試"}, ""
    
    try:
        # Create chat template
        messages = [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": text}
        ]
        
        prompt = tokenizer.apply_chat_template(
            messages, 
            tokenize=False, 
            add_generation_prompt=True
        )
        
        # Tokenize input
        inputs = tokenizer(prompt, return_tensors="pt", padding=True)
        
        # Generate response
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=64,
                temperature=0.1,
                pad_token_id=tokenizer.eos_token_id,
                eos_token_id=tokenizer.eos_token_id,
                do_sample=False,
            )
        
        # Extract assistant's response
        prompt_length = len(inputs.input_ids[0])
        assistant_output = tokenizer.decode(outputs[0][prompt_length:], skip_special_tokens=True)
        
        # Validate and parse JSON
        is_valid, parsed, message = validate_json(assistant_output)
        
        if is_valid:
            return parsed, assistant_output
        else:
            return {"error": message}, assistant_output
            
    except Exception as e:
        return {"error": f"Processing error / 處理時出錯: {str(e)}"}, ""

# Create Gradio interface
def create_interface():
    """Create the Gradio interface"""
    chinese_examples = [
        "你好，我想訂明天晚上7點的位子，四位成人，電話是0912-345-678",
        "週六下午三點，兩位，電話0987654321",
        "預約下週三中午12點半，5人用餐，聯絡電話0912345678",
        "我要訂位，3個人，今天下午6點"
    ]
    
    english_examples = [
        "Hello, I'd like to reserve a table for 4 people tomorrow at 7 PM, phone number is 0912-345-678",
        "Saturday 3 PM, 2 people, phone 0987654321",
        "Reservation for next Wednesday at 12:30 PM, 5 people, contact number 0912345678",
        "I want to make a reservation, 3 people, today at 6 PM"
    ]
    
    # Language-specific text dictionaries
    text_en = {
        "title": "🍽️ Restaurant Reservation Info Extractor",
        "description": "Use AI to automatically extract reservation information from messages",
        "input_label": "Input reservation message",
        "input_placeholder": "e.g., Hello, I'd like to reserve a table for 4 people tomorrow at 7 PM, phone number is 0912-345-678",
        "button_text": "Extract Information",
        "json_label": "Extracted Result",
        "raw_label": "Raw Output",
        "instructions_title": "ℹ️ Instructions",
        "instructions": """**Supported information:**
- 👥 Number of people (num_people)
- 📅 Reservation date/time (reservation_date)
- 📞 Phone number (phone_num)

**Notes:**
- First-time model loading may take a few minutes
- If you encounter errors, try refreshing the page
- The model outputs results in JSON format""",
        "footer": "Powered by [Together AI](https://together.ai) | Model: Luigi/gemma-3-270m-it-dinercall-ner",
        "examples_title": "Examples",
        "chinese_examples": "Chinese Examples",
        "english_examples": "English Examples"
    }
    
    text_zh = {
        "title": "🍽️ 餐廳訂位資訊提取器",
        "description": "使用AI從中文訊息中自動提取訂位資訊",
        "input_label": "輸入訂位訊息",
        "input_placeholder": "例如: 你好，我想訂明天晚上7點的位子，四位成人，電話是0912-345-678",
        "button_text": "提取資訊",
        "json_label": "提取結果",
        "raw_label": "原始輸出",
        "instructions_title": "ℹ️ 使用說明",
        "instructions": """**支援提取的資訊:**
- 👥 人數 (num_people)
- 📅 預訂日期/時間 (reservation_date)
- 📞 電話號碼 (phone_num)

**注意事項:**
- 首次加載模型可能需要幾分鐘時間
- 如果遇到錯誤，請嘗試刷新頁面
- 模型會輸出JSON格式的結果""",
        "footer": "由 [Together AI](https://together.ai) 提供技術支持 | 模型: Luigi/gemma-3-270m-it-dinercall-ner",
        "examples_title": "示例",
        "chinese_examples": "中文示例",
        "english_examples": "英文示例"
    }
    
    with gr.Blocks(
        title="Restaurant Reservation Info Extractor",
        theme=gr.themes.Soft()
    ) as demo:
        # Language selector
        language = gr.Radio(
            choices=["English", "中文"],
            value="English",
            label="Language / 語言",
            interactive=True
        )
        
        # Create components that will be updated based on language
        title_md = gr.Markdown("# 🍽️ Restaurant Reservation Info Extractor")
        description_md = gr.Markdown("Use AI to automatically extract reservation information from messages")
        
        with gr.Row():
            with gr.Column():
                input_text = gr.Textbox(
                    label="Input reservation message",
                    placeholder="e.g., Hello, I'd like to reserve a table for 4 people tomorrow at 7 PM, phone number is 0912-345-678",
                    lines=3
                )
                submit_btn = gr.Button("Extract Information", variant="primary")
                
                examples_title_md = gr.Markdown("### Examples")
                chinese_examples_title_md = gr.Markdown("### Chinese Examples")
                gr.Examples(
                    examples=chinese_examples,
                    inputs=input_text,
                    label="Chinese Examples"
                )
                
                english_examples_title_md = gr.Markdown("### English Examples")
                gr.Examples(
                    examples=english_examples,
                    inputs=input_text,
                    label="English Examples"
                )
                
            with gr.Column():
                json_output = gr.JSON(label="Extracted Result")
                raw_output = gr.Textbox(
                    label="Raw Output",
                    interactive=False,
                    lines=3
                )
        
        # Info panel - Create the Accordion but don't use it as an output
        with gr.Accordion("ℹ️ Instructions", open=False) as instructions_accordion:
            instructions_md = gr.Markdown("""**Supported information:**
- 👥 Number of people (num_people)
- 📅 Reservation date/time (reservation_date)
- 📞 Phone number (phone_num)

**Notes:**
- First-time model loading may take a few minutes
- If you encounter errors, try refreshing the page
- The model outputs results in JSON format""")
        
        # Footer
        footer_md = gr.Markdown("Powered by [Together AI](https://together.ai) | Model: Luigi/gemma-3-270m-it-dinercall-ner")
        
        # Function to update interface based on language selection
        def update_interface(language):
            texts = text_en if language == "English" else text_zh
            return [
                f"# {texts['title']}",  # title_md
                texts['description'],  # description_md
                gr.update(label=texts['input_label'], placeholder=texts['input_placeholder']),  # input_text
                texts['button_text'],  # submit_btn
                gr.update(label=texts['json_label']),  # json_output
                gr.update(label=texts['raw_label']),  # raw_output
                texts['instructions'],  # instructions_md
                texts['footer'],  # footer_md
                f"### {texts['examples_title']}",  # examples_title_md
                f"### {texts['chinese_examples']}",  # chinese_examples_title_md
                f"### {texts['english_examples']}"  # english_examples_title_md
            ]
        
        # Connect the function to the button
        submit_btn.click(
            fn=extract_reservation_info,
            inputs=input_text,
            outputs=[json_output, raw_output]
        )
        
        # Connect language selector to update interface - REMOVE ACCORDION FROM OUTPUTS
        language.change(
            fn=update_interface,
            inputs=language,
            outputs=[
                title_md,
                description_md,
                input_text,
                submit_btn,
                json_output,
                raw_output,
                instructions_md,        # Update only the Markdown inside Accordion
                footer_md,
                examples_title_md,
                chinese_examples_title_md,
                english_examples_title_md
            ]
        )
    
    return demo

# Create and launch the interface
if __name__ == "__main__":
    # Pre-load the model when the app starts
    load_model()
    
    demo = create_interface()
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )