catalog-agent 0.1.5

AI-powered catalog agent for product discovery and recommendations

Catalog Agent

AI-powered catalog agent for product discovery and recommendations using LangChain, Pydantic, and Supabase.

Features

• Intelligent Product Search: AI-powered semantic search and filtering
• Intent Detection: Advanced intent recognition with synonym matching
• Conversation Management: Multi-session conversation tracking
• User Preferences: Personalized recommendations based on user preferences
• LangChain Integration: Built on LangChain for robust AI agent capabilities
• Type Safety: Full type safety with Pydantic models
• Supabase Integration: Seamless integration with Supabase for product data

Installation

From PyPI (Recommended)

pip install catalog-agent

From Source

git clone https://github.com/komatadi/catalog-agent.git
cd catalog-agent
pip install -e .

Quick Start

Basic Usage

from catalog_agent import CatalogAgent, AgentConfig

# Initialize the agent
config = AgentConfig(
    openai_api_key="your_openai_api_key",
    supabase_functions_url="your_supabase_functions_url",
    gpt_actions_api_key="your_gpt_actions_api_key"
)

agent = CatalogAgent(config)

# Chat with the agent
response = agent.chat("I need a red dress for a wedding")
print(response.message)

# Access product results
if response.products:
    for product in response.products:
        print(f"- {product.title}: {product.url}")

Environment Variables

Set the following environment variables:

export OPENAI_API_KEY="your_openai_api_key"
export SUPABASE_FUNCTIONS_URL="your_supabase_functions_url"
export GPT_ACTIONS_API_KEY="your_gpt_actions_api_key"

Or create a .env file:

# .env
OPENAI_API_KEY=your_openai_api_key_here
SUPABASE_FUNCTIONS_URL=your_supabase_functions_url_here
GPT_ACTIONS_API_KEY=your_gpt_actions_api_key_here

Configuration

Agent Configuration

from catalog_agent import AgentConfig

config = AgentConfig(
    openai_api_key="your_key",           # Required
    supabase_functions_url="your_url",   # Required
    gpt_actions_api_key="your_key",      # Required
    model="gpt-4o-mini",                 # Optional, default: "gpt-4"
    temperature=0.1,                     # Optional, default: 0.1
    max_tokens=2000,                     # Optional, default: 2000
    session_id="user_123",               # Optional
    debug=False,                         # Optional, default: False
    max_iterations=3,                    # Optional, default: 3
    max_execution_time=30,               # Optional, default: 30
    early_stopping_method="generate",    # Optional, default: "generate"
    log_level="INFO",                    # Optional, default: "INFO"
    enable_agent_verbose=False           # Optional, default: None
)

Performance & Configuration

Model Selection

For optimal performance and cost, we recommend:

• gpt-4o-mini: Best balance of cost, speed, and capability (~90% cheaper than GPT-4)
• gpt-3.5-turbo: Lowest cost, good for simple queries
• gpt-4-turbo: Higher cost but better for complex reasoning

Agent Execution Limits

Control agent behavior with these settings:

• max_iterations=3: Limits reasoning loops (prevents infinite loops)
• max_execution_time=30: Timeout in seconds (prevents runaway execution)
• early_stopping_method="generate": Returns best available response when hitting limits

Logging Configuration

• log_level="INFO": Production logging (shows important events)
• log_level="DEBUG": Development logging (shows all internal processing)
• enable_agent_verbose=False: Suppresses LangChain's step-by-step reasoning
• enable_agent_verbose=True: Shows tool calls and agent decision-making

Environment Variables

Set these in your .env file:

# Model Configuration
OPENAI_MODEL=gpt-4o-mini

# Agent Execution Limits
MAX_AGENT_ITERATIONS=3
MAX_EXECUTION_TIME=30
EARLY_STOPPING_METHOD=generate

# Logging Configuration
LOG_LEVEL=INFO
ENABLE_AGENT_VERBOSE=false

Execution Modes

The agent supports two execution modes:

Direct Mode (Default, Recommended)

Fast, predictable, and cost-effective for product discovery:

config = AgentConfig(
    # ... other config
    use_direct_mode=True,           # Default
    use_llm_formatting=False        # Optional natural language formatting
)

Characteristics:

• Token usage: < 2K tokens per request
• Response time: < 500ms
• Cost: ~$0.001 per request with gpt-4o-mini
• Behavior: Deterministic tool calling based on intent

Best for: Product search, filtering, recommendations

AgentExecutor Mode

LangChain agentic reasoning for complex scenarios:

config = AgentConfig(
    # ... other config
    use_direct_mode=False,
    max_iterations=3,
    max_execution_time=30,
    enable_agent_verbose=False
)

Characteristics:

• Token usage: 5K-10K tokens per request (with limits)
• Response time: 2-5 seconds
• Cost: ~$0.01-0.02 per request with gpt-4o-mini
• Behavior: Agent decides which tools to use and can loop

Best for: Future features like emotion detection, complex multi-step reasoning

Environment Variables

# Execution Mode
USE_DIRECT_MODE=true              # Use direct mode (recommended)

# Direct Mode Options
USE_LLM_FORMATTING=false          # Natural language formatting

# AgentExecutor Mode Options (when USE_DIRECT_MODE=false)
MAX_AGENT_ITERATIONS=3
MAX_EXECUTION_TIME=30
ENABLE_AGENT_VERBOSE=false

Configuration Files

The agent uses configuration files in the config/ directory:

• instructions.yaml: Agent instructions and prompts
• actions.yaml: Available actions and tools
• intent-synonyms.json: Intent detection synonyms
• DiscoverProducts.json: Product discovery configuration
• tool-playbook.md: Tool usage guidelines

Advanced Usage

Multi-User Chatbot Integration

class MyChatbot:
    def __init__(self):
        self.agent = CatalogAgent(config)
        self.user_sessions = {}
    
    def process_message(self, user_id, message):
        session_id = self.user_sessions.get(user_id, f"user_{user_id}")
        response = self.agent.chat(message, session_id)
        return response
    
    def update_preferences(self, user_id, preferences):
        session_id = self.user_sessions.get(user_id)
        if session_id:
            self.agent.update_user_preferences(preferences, session_id)

User Preferences

# Update user preferences
agent.update_user_preferences({
    "sizes": ["M", "L"],
    "colors": ["red", "blue"],
    "brands": ["Nike", "Adidas"],
    "occasions": ["casual", "formal"]
}, session_id)

# Get conversation context
context = agent.get_conversation_context(session_id)
print(f"User preferences: {context.user_preferences}")

Session Management

# Reset conversation
agent.reset_conversation(session_id)

# Get conversation context
context = agent.get_conversation_context(session_id)

# Check agent health
is_healthy = agent.health_check()

Examples

Interactive Chat

Run the interactive chat example:

python examples/simple_chat.py

Chatbot Integration

See the chatbot integration example:

python examples/chatbot_integration.py

Testing

Smoke Test

Run the complete smoke test:

python tests/smoke_test.py

Intent Service Test

Test intent detection in isolation:

python tests/intent_smoke.py

Configuration Health Check

Validate configuration files:

python tests/config_health.py

Response Structure

The catalog-agent returns structured responses that are designed for easy chatbot integration. Understanding these response fields is essential for proper integration.

AgentResponse

The main response object returned by agent.chat():

{
    "message": str,                    # Required: Human-readable response text
    "products": List[ProductResult],   # Optional: Array of found products
    "success": bool,                   # Required: Whether the response was successful
    "metadata": Dict[str, Any]         # Optional: Additional response metadata
}

Core Fields

• message (str, required): The main response text that should be displayed to the user
• products (List[ProductResult], optional): Array of product results when products are found
• success (bool, default=True): Indicates whether the response was successful
• metadata (Dict[str, Any], optional): Additional response metadata

ProductResult

When products are found, each product in the products array contains:

{
    "handle": str,           # Required: Product handle/identifier
    "title": str,            # Required: Product title/name
    "url": str,              # Required: Direct product URL
    "image_url": str,        # Optional: Product image URL
    "score": float,          # Optional: Relevance score (0.0-1.0)
    "boosted": bool          # Optional: Whether the product is boosted/promoted
}

Product Fields

• handle (str, required): Product handle/identifier
• title (str, required): Product title/name
• url (str, required): Direct product URL
• image_url (str, optional): Product image URL
• score (float, optional): Search relevance score (0.0-1.0)
• boosted (bool, optional): Whether the product is boosted/promoted

Metadata Fields

The metadata object typically includes:

• session_id (str): Session identifier for conversation tracking
• duration_ms (float): Response time in milliseconds
• workflow (str): Workflow type used (e.g., "direct_mode", "semantic_fallback", "intent_clarification")
• tools_used (List[str], optional): List of tools used during processing
• mode (str, optional): Execution mode ("direct" or "agent_executor")
• filters_applied (Dict, optional): Applied search filters
• products_found (int, optional): Number of products found
• error (str, optional): Error message if something went wrong

Response Examples

Successful Product Search

{
    "message": "Here are the top matches for 'red dress':\n1. Red Evening Dress — https://shop.styledgenie.com/products/red-evening-dress\n2. Summer Red Dress — https://shop.styledgenie.com/products/summer-red-dress",
    "products": [
        {
            "handle": "red-evening-dress",
            "title": "Red Evening Dress",
            "url": "https://shop.styledgenie.com/products/red-evening-dress",
            "image_url": "https://example.com/image.jpg",
            "score": 0.95,
            "boosted": false
        }
    ],
    "success": true,
    "metadata": {
        "session_id": "chat-1234567890",
        "duration_ms": 450.5,
        "workflow": "direct_mode",
        "tools_used": ["search_products"],
        "products_found": 5
    }
}

Intent Clarification Needed

{
    "message": "I'm close, but I need a bit more detail to lock onto the right products...",
    "products": null,
    "success": false,
    "metadata": {
        "session_id": "chat-1234567890",
        "duration_ms": 200.3,
        "workflow": "intent_clarification",
        "tools_used": []
    }
}

Error Response

{
    "message": "I ran into an unexpected issue reaching the catalog. Please try again in a moment.",
    "products": null,
    "success": false,
    "metadata": {
        "session_id": "chat-1234567890",
        "duration_ms": 1500.0,
        "workflow": "error"
    }
}

Integration Example

from catalog_agent import CatalogAgent, AgentConfig

# Initialize agent
config = AgentConfig(
    openai_api_key="your_key",
    supabase_functions_url="your_url", 
    gpt_actions_api_key="your_key"
)
agent = CatalogAgent(config)

# Process user message
response = agent.chat("I need a red dress for a wedding", session_id="user123")

# Access response fields
print(f"Message: {response.message}")
print(f"Success: {response.success}")

if response.products:
    for product in response.products:
        print(f"Product: {product.title}")
        print(f"URL: {product.url}")
        print(f"Score: {product.score}")

# Access metadata
print(f"Response time: {response.metadata.get('duration_ms', 0):.0f}ms")
print(f"Tools used: {response.metadata.get('tools_used', [])}")

API Reference

CatalogAgent

Main agent class for product discovery and recommendations.

Methods

• chat(message: str, session_id: Optional[str] = None) -> AgentResponse
• stream_chat(message: str, session_id: Optional[str] = None) -> Iterator[str]
• reset_conversation(session_id: Optional[str] = None) -> None
• get_conversation_context(session_id: Optional[str] = None) -> Optional[ConversationContext]
• update_user_preferences(preferences: Dict[str, Any], session_id: Optional[str] = None) -> None
• health_check() -> bool

Development

Setup Development Environment

git clone https://github.com/komatadi/catalog-agent.git
cd catalog-agent
pip install -e ".[dev]"

Running Tests

# Run all tests
pytest

# Run specific test
pytest tests/smoke_test.py

# Run with coverage
pytest --cov=catalog_agent

Code Quality

# Format code
black catalog_agent/

# Lint code
ruff catalog_agent/

# Type checking
mypy catalog_agent/

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests for new functionality
5. Ensure all tests pass
6. Submit a pull request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Support

For support and questions:

• Create an issue on GitHub
• Check the examples in the examples/ directory
• Review the test files for usage patterns

Changelog

0.1.5

• BREAKING: Removed config_path parameter from AgentConfig - now uses built-in config files
• Agent automatically loads configuration from package's ./config directory
• Eliminates need for external config files and simplifies deployment
• Updated all examples and test scripts to remove config_path parameter
• Enhanced error handling for missing built-in configuration files
• Improved package reliability and reduced setup complexity

0.1.4

• Added Dual Mode Architecture: Direct Mode (default) and AgentExecutor Mode
• Direct Mode: 95% cost reduction, < 2K tokens, < 500ms response time
• Fixed bug: max_iterations=0 now works correctly in AgentExecutor mode
• Fixed bug: Verbose logging properly suppressed when disabled
• Added USE_DIRECT_MODE and USE_LLM_FORMATTING configuration options
• Enhanced config health check to validate environment variables
• Updated documentation with execution mode comparison and best practices
• AgentExecutor mode preserved for future complex reasoning features

0.1.3

• Added configurable agent execution limits (max_iterations, max_execution_time, early_stopping_method)
• Added structured logging with configurable log levels (LOG_LEVEL)
• Added ENABLE_AGENT_VERBOSE flag for controlling LangChain's internal logging
• Optimized system prompt to reduce token usage (~70% reduction)
• Added support for production-grade logging control
• Performance improvements and token budget management
• Recommended model: gpt-4o-mini or gpt-3.5-turbo for cost/speed optimization
• Fixed token overflow issue causing rate limit errors with GPT-4

0.1.2

• BREAKING: Relaxed Pydantic dependency constraints for better compatibility
• Changed pydantic==2.9.2 to pydantic>=2.9.2,<3.0.0
• Changed pydantic-settings==2.6.1 to pydantic-settings>=2.6.1,<3.0.0
• Resolves dependency conflicts with openai-chatkit and openai-agents packages
• Maintains backward compatibility with Pydantic 2.9.2+
• All existing functionality preserved with improved package compatibility

0.1.1

• Bug fixes and improvements
• Enhanced error handling
• Better session management

0.1.0

• Initial release
• Basic agent functionality
• Intent detection and synonym matching
• Product search and filtering
• Multi-session conversation management
• User preference tracking
• Comprehensive examples and tests