app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import warnings
warnings.filterwarnings("ignore")

class LlamaAddressCompletion:
    def __init__(self):
        self.model_name = "shiprocket-ai/open-llama-1b-address-completion"
        self.model = None
        self.tokenizer = None
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.load_model()
    
    def load_model(self):
        """Load the Llama model and tokenizer"""
        try:
            print("Loading Llama 3.2-1B Address Completion model...")
            self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
            
            # Load model with appropriate settings for the space
            self.model = AutoModelForCausalLM.from_pretrained(
                self.model_name,
                torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
                device_map="auto" if torch.cuda.is_available() else None,
                trust_remote_code=True
            )
            
            if not torch.cuda.is_available():
                self.model = self.model.to(self.device)
            
            self.model.eval()
            print("✅ Model loaded successfully!")
            
        except Exception as e:
            print(f"❌ Error loading model: {str(e)}")
            raise e
    
    def extract_address_components(self, address, max_new_tokens=150):
        """Extract address components using the model"""
        if not address.strip():
            return "Please provide an address to extract components from."
        
        try:
            # Format prompt for Llama 3.2-1B-Instruct
            prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>

Extract address components from: {address}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""
            
            # Tokenize
            inputs = self.tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
            
            # Move inputs to the same device as the model
            device = next(self.model.parameters()).device
            inputs = {k: v.to(device) for k, v in inputs.items()}
            
            # Generate
            with torch.no_grad():
                outputs = self.model.generate(
                    **inputs,
                    max_new_tokens=max_new_tokens,
                    temperature=0.1,
                    top_p=0.9,
                    do_sample=True,
                    pad_token_id=self.tokenizer.eos_token_id,
                    repetition_penalty=1.05
                )
            
            # Decode only the new tokens
            input_length = inputs['input_ids'].shape[1]
            generated_tokens = outputs[0][input_length:]
            response = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
            
            return response.strip()
            
        except Exception as e:
            return f"Error processing address: {str(e)}"
    
    def complete_partial_address(self, partial_address, max_new_tokens=100):
        """Complete a partial address"""
        if not partial_address.strip():
            return "Please provide a partial address to complete."
        
        try:
            # Format prompt for address completion
            prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>

Complete this partial address: {partial_address}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""
            
            # Tokenize
            inputs = self.tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
            
            # Move inputs to the same device as the model
            device = next(self.model.parameters()).device
            inputs = {k: v.to(device) for k, v in inputs.items()}
            
            # Generate
            with torch.no_grad():
                outputs = self.model.generate(
                    **inputs,
                    max_new_tokens=max_new_tokens,
                    temperature=0.2,
                    top_p=0.9,
                    do_sample=True,
                    pad_token_id=self.tokenizer.eos_token_id,
                    repetition_penalty=1.05
                )
            
            # Decode only the new tokens
            input_length = inputs['input_ids'].shape[1]
            generated_tokens = outputs[0][input_length:]
            response = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
            
            return response.strip()
            
        except Exception as e:
            return f"Error completing address: {str(e)}"
    
    def standardize_address(self, address, max_new_tokens=150):
        """Standardize an address format"""
        if not address.strip():
            return "Please provide an address to standardize."
        
        try:
            # Format prompt for address standardization
            prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>

Standardize this address into proper format: {address}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""
            
            # Tokenize
            inputs = self.tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
            
            # Move inputs to the same device as the model
            device = next(self.model.parameters()).device
            inputs = {k: v.to(device) for k, v in inputs.items()}
            
            # Generate
            with torch.no_grad():
                outputs = self.model.generate(
                    **inputs,
                    max_new_tokens=max_new_tokens,
                    temperature=0.1,
                    top_p=0.9,
                    do_sample=True,
                    pad_token_id=self.tokenizer.eos_token_id,
                    repetition_penalty=1.05
                )
            
            # Decode only the new tokens
            input_length = inputs['input_ids'].shape[1]
            generated_tokens = outputs[0][input_length:]
            response = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
            
            return response.strip()
            
        except Exception as e:
            return f"Error standardizing address: {str(e)}"

# Initialize the model
print("Initializing Llama Address Completion system...")
try:
    llama_system = LlamaAddressCompletion()
    print("System ready!")
except Exception as e:
    print(f"Failed to initialize system: {e}")
    llama_system = None

def extract_components_interface(address_text):
    """Interface function for component extraction"""
    if llama_system is None:
        return "❌ Model not loaded. Please check the logs."
    
    result = llama_system.extract_address_components(address_text)
    return f"**Input:** {address_text}\n\n**Extracted Components:**\n{result}"

def complete_address_interface(partial_address):
    """Interface function for address completion"""
    if llama_system is None:
        return "❌ Model not loaded. Please check the logs."
    
    result = llama_system.complete_partial_address(partial_address)
    return f"**Partial Address:** {partial_address}\n\n**Completed Address:**\n{result}\n\n*⚠️ Note: This feature has limited training data and results may vary in quality.*"

def standardize_address_interface(address_text):
    """Interface function for address standardization"""
    if llama_system is None:
        return "❌ Model not loaded. Please check the logs."
    
    result = llama_system.standardize_address(address_text)
    return f"**Original:** {address_text}\n\n**Standardized:**\n{result}\n\n*⚠️ Note: This feature has limited training data and results may vary in quality.*"

# Sample data
sample_addresses = [
    "C-704, Gayatri Shivam, Thakur Complex, Kandivali East, 400101",
    "Villa 141, Geown Oasis, V Kallahalli, Off Sarjapur, Bengaluru, Karnataka, 562125",
    "E401 Supertech Icon Indrapam 201301 UP",
    "Shop No 123, Sunshine Apartments, Andheri West, Mumbai, 400058",
    "Flat 201, MG Road, Bangalore, Karnataka, 560001"
]

partial_addresses = [
    "C-704, Gayatri Shivam, Thakur Complex",
    "Villa 141, Geown Oasis, V Kallahalli",
    "E401 Supertech Icon",
    "Shop No 123, Sunshine Apartments",
    "Flat 201, MG Road, Bangalore"
]

informal_addresses = [
    "c704 gayatri shivam thakur complex kandivali e 400101",
    "villa141 geown oasis vkallahalli off sarjapur blr kar 562125",
    "e401 supertech icon indrapam up 201301",
    "shop123 sunshine apts andheri w mumbai 400058"
]

# Create Gradio interface
with gr.Blocks(title="Llama Address Intelligence", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🦙 Llama 3.2-1B Address Intelligence
    
    Powered by a fine-tuned Llama 3.2-1B model specialized for Indian address processing. 
    
    **⭐ Best Performance**: Entity extraction from complete addresses  
    **⚠️ Limited Performance**: Address completion and standardization (limited training data)
    
    **Model:** [shiprocket-ai/open-llama-1b-address-completion](https://huggingface.co/shiprocket-ai/open-llama-1b-address-completion)
    """)
    
    with gr.Tab("📋 Extract Components"):
        gr.Markdown("⭐ **BEST PERFORMANCE** - Extract structured components from complete addresses")
        with gr.Row():
            with gr.Column(scale=1):
                extract_input = gr.Textbox(
                    label="Enter Address",
                    placeholder="e.g., C-704, Gayatri Shivam, Thakur Complex, Kandivali East, 400101",
                    lines=3
                )
                extract_btn = gr.Button("🔍 Extract Components", variant="primary")
                
                gr.Markdown("### Sample Addresses:")
                extract_samples = []
                for addr in sample_addresses:
                    btn = gr.Button(addr, size="sm")
                    btn.click(fn=lambda x=addr: x, outputs=extract_input)
                    extract_samples.append(btn)
            
            with gr.Column(scale=1):
                extract_output = gr.Markdown(
                    value="Enter an address and click 'Extract Components' to see structured breakdown."
                )
        
        extract_btn.click(
            fn=extract_components_interface,
            inputs=extract_input,
            outputs=extract_output
        )
        
        extract_input.submit(
            fn=extract_components_interface,
            inputs=extract_input,
            outputs=extract_output
        )
    
    with gr.Tab("✨ Complete Address"):
        gr.Markdown("⚠️ **EXPERIMENTAL** - Complete partial addresses (limited training data - results may vary)")
        with gr.Row():
            with gr.Column(scale=1):
                complete_input = gr.Textbox(
                    label="Enter Partial Address",
                    placeholder="e.g., C-704, Gayatri Shivam, Thakur Complex",
                    lines=3
                )
                complete_btn = gr.Button("🚀 Complete Address", variant="primary")
                
                gr.Markdown("### Sample Partial Addresses:")
                complete_samples = []
                for addr in partial_addresses:
                    btn = gr.Button(addr, size="sm")
                    btn.click(fn=lambda x=addr: x, outputs=complete_input)
                    complete_samples.append(btn)
            
            with gr.Column(scale=1):
                complete_output = gr.Markdown(
                    value="Enter a partial address and click 'Complete Address' to see the AI completion."
                )
        
        complete_btn.click(
            fn=complete_address_interface,
            inputs=complete_input,
            outputs=complete_output
        )
        
        complete_input.submit(
            fn=complete_address_interface,
            inputs=complete_input,
            outputs=complete_output
        )
    
    with gr.Tab("📐 Standardize Format"):
        gr.Markdown("⚠️ **EXPERIMENTAL** - Convert informal addresses to standardized format (limited training data - results may vary)")
        with gr.Row():
            with gr.Column(scale=1):
                standardize_input = gr.Textbox(
                    label="Enter Informal Address",
                    placeholder="e.g., c704 gayatri shivam thakur complex kandivali e 400101",
                    lines=3
                )
                standardize_btn = gr.Button("📏 Standardize Format", variant="primary")
                
                gr.Markdown("### Sample Informal Addresses:")
                standardize_samples = []
                for addr in informal_addresses:
                    btn = gr.Button(addr, size="sm")
                    btn.click(fn=lambda x=addr: x, outputs=standardize_input)
                    standardize_samples.append(btn)
            
            with gr.Column(scale=1):
                standardize_output = gr.Markdown(
                    value="Enter an informal address and click 'Standardize Format' to see the cleaned version."
                )
        
        standardize_btn.click(
            fn=standardize_address_interface,
            inputs=standardize_input,
            outputs=standardize_output
        )
        
        standardize_input.submit(
            fn=standardize_address_interface,
            inputs=standardize_input,
            outputs=standardize_output
        )
    
    with gr.Tab("ℹ️ Model Information"):
        gr.Markdown("""
        ## 🦙 About Llama 3.2-1B Address Completion
        
        ### Model Specifications
        - **Base Model**: meta-llama/Llama-3.2-1B-Instruct
        - **Parameters**: 1.24B parameters
        - **Model Size**: ~2.47GB
        - **Architecture**: Causal Language Model (Autoregressive)
        - **Max Context**: 131,072 tokens
        - **Precision**: FP16 for GPU, FP32 for CPU
        
        ### Key Features
        - **Lightweight**: Only 1B parameters for fast inference
        - **Specialized**: Fine-tuned specifically for Indian addresses
        - **Versatile**: Handles extraction, completion, and standardization
        - **Efficient**: Optimized for real-time applications
        - **Context-Aware**: Understands relationships between address components
        
        ### Supported Address Components
        - **Building Names**: Apartments, complexes, towers, malls
        - **Localities**: Areas, neighborhoods, sectors  
        - **Pincodes**: 6-digit Indian postal codes
        - **Cities**: Major and minor Indian cities
        - **States**: All Indian states and union territories
        - **Sub-localities**: Sectors, phases, blocks
        - **Road Names**: Streets, lanes, main roads
        - **Landmarks**: Notable reference points
        
        ### Performance Notes
        - **⭐ Entity Extraction**: Excellent performance - primary use case
        - **⚠️ Address Completion**: Limited training data - experimental feature
        - **⚠️ Format Standardization**: Limited training data - experimental feature
        
        **Recommendation**: Use this model primarily for address component extraction.
        
        ### Use Cases
        - **E-commerce**: Auto-complete checkout addresses
        - **Forms**: Intelligent address suggestions
        - **Data Cleaning**: Standardize legacy address databases
        - **Mobile Apps**: On-device address processing
        - **APIs**: Real-time address validation services
        
        ### Performance Tips
        - Use lower temperatures (0.1-0.3) for factual outputs
        - Keep prompts under 512 tokens for optimal speed
        - Process in batches for high-throughput scenarios
        - Works best with Llama chat format prompts
        """)
    
    gr.Markdown("""
    ---
    **Powered by:** [Llama 3.2-1B Address Completion](https://huggingface.co/shiprocket-ai/open-llama-1b-address-completion) | 
    **License:** Apache 2.0 | 
    **Developed by:** Shiprocket AI Team
    
    This model demonstrates the power of lightweight LLMs for specialized address intelligence tasks.
    """)

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