Building an Intelligent E-commerce Assistant: LLM-Powered Intent Recognition Using Cohere and Gradio

Eddie Otudor — Fri, 08 Nov 2024 13:24:00 +0000

As e-commerce continues to evolve, understanding customer intent has become crucial for providing exceptional shopping experiences. In this article, I'll share how we built a sophisticated intent recognition system using Cohere's Command model and Gradio, creating an interactive demo that showcases the power of LLMs in e-commerce.

The Challenge

Traditional e-commerce search and navigation systems often struggle with understanding natural language queries like "I need a waterproof watch that won't break the bank" or "Can you help me find a gift for my tech-savvy dad?". These systems typically rely on keyword matching and predefined rules, missing the nuanced understanding that modern customers expect.

Solution: LLM-Powered Intent Recognition

We've built an interactive demo using Gradio that showcases how Large Language Models (LLMs) can revolutionize e-commerce interactions. Our system:

Understands complex, natural language queries
Extracts relevant entities and specifications
Determines user intent with high accuracy
Provides detailed analysis in real-time

Technical Implementation

1. Core Intent Recognition System

@dataclass
class IntentResponse:
    intent: str
    confidence: float
    entities: Dict
    suggested_action: str
    explanation: str

class EcommerceLLMIntentRecognizer:
    def __init__(self):
        api_key = os.getenv('COHERE_API_KEY')
        self.co = cohere.Client(api_key)

        self.valid_intents = {
            'product_search': 'SEARCH_CATALOG',
            'price_inquiry': 'FETCH_PRICE',
            'order_status': 'CHECK_ORDER_STATUS',
            'return_request': 'INITIATE_RETURN',
            'cart_management': 'MODIFY_CART',
            'availability_check': 'CHECK_INVENTORY',
            'checkout_help': 'ASSIST_CHECKOUT',
            'shipping_info': 'PROVIDE_SHIPPING_INFO'
        }

2. LLM Prompt Engineering

The key to accurate intent recognition lies in our carefully crafted prompt:

def _generate_prompt(self, query: str) -> str:
    return f"""As an e-commerce AI assistant, analyze the following customer query and extract the shopping intent, relevant entities, and determine the appropriate action.

Valid intents are: {', '.join(self.valid_intents.keys())}

Customer Query: "{query}"

Provide your analysis in the following JSON format:
{{
    "intent": "the_identified_intent",
    "confidence": 0.XX,
    "entities": {{
        "product": "identified_product",
        "category": "product_category",
        "specifications": ["any", "relevant", "specs"],
        "price_range": {{
            "min": "if_mentioned",
            "max": "if_mentioned"
        }}
    }},
    "explanation": "Brief explanation of why this intent was chosen"
}}"""

3. Interactive Demo with Gradio

One of the most powerful aspects of our implementation is the interactive demo built with Gradio:

def process_query(user_query: str) -> str:
    try:
        recognizer = EcommerceLLMIntentRecognizer()
        response = recognizer.recognize_intent(user_query)

        return json.dumps({
            'timestamp': datetime.now().isoformat(),
            'query': user_query,
            'intent': response.intent,
            'confidence': response.confidence,
            'entities': response.entities,
            'suggested_action': response.suggested_action,
            'explanation': response.explanation
        }, indent=2)
    except ValueError as e:
        return json.dumps({
            'error': str(e),
            'hint': 'Please ensure COHERE_API_KEY is set in your .env file'
        }, indent=2)

# Create Gradio interface
iface = gr.Interface(
    fn=process_query,
    inputs=gr.Textbox(label="Enter customer query"),
    outputs=gr.JSON(label="Intent Analysis"),
    title="E-commerce LLM Intent Recognition System",
    description="Enter your query to see the detailed intent analysis.",
    examples=[
        ["I'm looking for a waterproof smart watch under $300"],
        ["Can you compare the iPhone 13 and iPhone 14 Pro?"],
        ["Need to return my order #ABC123, it's the wrong size"],
        ["Do you have this dress in size medium and in red?"],
        ["What's your shipping time to California?"]
    ]
)

Real-World Examples

Let's look at some example queries and their analysis:

1. Product Search with Specifications

Query: "I need a waterproof smartwatch under $300"

Response:

{
  "intent": "product_search",
  "confidence": 0.95,
  "entities": {
    "product": "smartwatch",
    "specifications": ["waterproof"],
    "price_range": {
      "max": 300
    }
  },
  "explanation": "Query specifies a product type with specific features and price constraint"
}

2. Order Status Check

Query: "Where's my order #ABC123?"

Response:

{
  "intent": "order_status",
  "confidence": 0.98,
  "entities": {
    "order_id": "ABC123"
  },
  "explanation": "Direct question about order status with order number"
}

Benefits of Our Approach

1. Interactive Demo

Instant feedback through Gradio interface
Easy to test different queries
Visual JSON output for clear understanding

2. Developer-Friendly

Simple setup process
Clear code structure
Easy to extend and modify

3. Business Value

Better customer understanding
Improved search accuracy
Automated query classification

Implementation Guide

Want to try it yourself? Here's how:

1. Install Requirements

pip install cohere gradio python-dotenv

2. Set Up Environment

Create a .env file:

COHERE_API_KEY=your_api_key_here

3. Run the Demo

python app.py

Future Enhancements

1. Enhanced Analysis

Sentiment analysis
Multiple intent detection
Confidence scoring improvements

2. UI Improvements

Visual intent mapping
Real-time suggestions
Mobile optimization

3. Integration Capabilities

API endpoint creation
Webhook support
Database integration

Conclusion

By combining the power of Cohere's LLM with Gradio's user-friendly interface, we've created a powerful tool for understanding e-commerce customer intentions. This demo showcases how modern AI can transform the way we interact with e-commerce platforms.

The application is running as a Hugging Face Space over here. Feel free to reach out if you have questions or want to discuss implementing similar solutions for your e-commerce platform.

Author's Note: I'm passionate about building practical AI solutions that bridge the gap between advanced technology and real-world applications. Connect with me to discuss more about AI in e-commerce or to explore collaboration opportunities.

DEV Community: Eddie Otudor