Let's build an AI agent using Laravel | Laravel AI SDK

Introduction to Laravel AI SDK

Overview of the Video

• The presenter, Aritya, introduces the video focused on using the newly released Laravel AI SDK to create a sophisticated sales agent.

• The main goal is to implement a chatbot that can respond to user queries about products and orders.

Project Setup

• A Laravel project with React and Inertia is used as a starter kit for simplicity, emphasizing the implementation of the Laravel AI SDK.

• The presenter mentions having models for categories, orders, and products, with seeded values in the database for demonstration.

Installing and Configuring the SDK

Installation Steps

• The installation process begins by copying a command to install the SDK via terminal.

• After installation, files including stubs and migrations need to be published using another command.

Database Migration

• Two tables are created: agent_conversations for storing conversation history and agent_conversation_messages for individual messages within those conversations.

• The presenter runs migration commands in terminal to set up these tables in their database.

Configuring API Keys and Models

Configuration Details

• Default settings use OpenAI APIs for text generation; other models like Gemini (for images) and Coher (for reranking) are also mentioned.

• Users can change default models or specify different ones based on specific tasks by updating their environment variables.

Setting Up API Keys

• The presenter explains how to add API keys into the .env file, specifically mentioning an OpenAI key setup.

Creating the Sales Assistant Agent

Defining Agent Characteristics

• An agent named "sales assistant" is created with capabilities to provide structured outputs based on user queries regarding products and orders.

Instructions for Agent Functionality

• Boundaries are defined so that the agent has access only to relevant data such as products, categories, and order information.

Implementing Tools for Enhanced Functionality

Tool Creation Process

• Instructions are provided on defining what tasks or tools will assist the agent in responding accurately to user inquiries about orders or product listings.

Utilizing Built-in Tools

• Discussion includes built-in tools available from various models like web search functionalities which can enhance responses based on user input.

Implementing User-Specific Tools in a Product Management System

Overview of Tool Integration

• The discussion begins with the introduction of various tools: product tool, list category tool, and list order tool for users. These tools are designed to cater to specific user needs.

• The speaker outlines the process of adding these tools into the system, emphasizing the importance of passing details to each tool for better functionality.

Setting Up Routes and Controllers

• A route is created in web.php to facilitate communication with an agent through a POST request named invoke agent.

• The speaker demonstrates how to create an AgentController using PHP Artisan command, which will handle requests related to invoking agents.

Validating Requests

• Before invoking the agent, it’s crucial to validate that a prompt has been provided by the user. This involves setting validation rules for incoming requests.

• The method used can either pass user information directly or utilize a prompt function; flexibility is highlighted as key in this design.

Handling Agent Responses

• A response from the sales assistant is generated based on validated input. If structured output is needed later, adjustments may be required.

• The constructor within the sales assistant captures user data, ensuring that interactions are personalized and relevant.

Developing Product Tool Functionality

• Focus shifts towards creating functionality for listing products. Options include displaying all products or filtering them based on category ID.

• Validation checks are implemented for category inputs; if no category ID is provided, all products will be listed by default.

Querying Products Based on Categories

• A query mechanism is established where products can be filtered based on their associated categories using Laravel's Eloquent ORM capabilities.

• The implementation includes checking relationships between products and categories before returning results back to users.

Frontend Interaction with React

• Transitioning to frontend development using React, state management for messages from users is discussed alongside handling input values effectively.

• An event handler (handleSend) is set up for sending messages via fetch calls; this includes preparing JSON data for transmission.

Testing API Calls and Handling CSRF Tokens

• Initial testing reveals issues due to missing CSRF tokens in fetch requests; solutions involve adding meta tags and configuring headers correctly.

Product Management and User Interaction in Development

Parsing JSON and Displaying Products

• The initial error encountered was related to parsing data as JSON. The solution involved changing the parsing method to text, allowing for a successful display of all products.

• A state management approach using useState is introduced to handle chat messages, with an optional ID for each message since it's not always necessary.

• Responses are updated by appending new user messages along with an incremented ID, ensuring that the assistant's responses are tracked effectively.

User Message Handling and UI Improvements

• Before sending a message, it’s added to the responses list as a user message, enhancing interaction tracking.

• After sending a request (e.g., "show me all products"), there’s a need to clear the input field and potentially show a loading state while fetching data from the database.

Category Filtering and Product Relations

• Users can filter products by category (e.g., men's tops), demonstrating how specific queries yield relevant product results based on category relations in the database.

• The implementation of a markdown library improves visual presentation by converting markdown text into HTML safely using DOM Purify to prevent XSS attacks.

Order Management and User-Specific Tools

• New tools have been added for listing categories based on provided names, alongside an order tool that retrieves orders linked to specific users via their IDs.

• Emphasis is placed on maintaining clean input schemas when passing user IDs directly into functions or classes for order retrieval.

Conversation Tracking and Database Integration

• The system allows saving conversations automatically through Laravel AI SDK integration, eliminating manual message handling.

• When creating or interacting with an assistant instance, either a new conversation ID can be generated or an existing one can be passed to retrieve past messages efficiently.

Testing Functionality and Final Demonstration

• A demonstration shows how user interactions (like asking for product lists or orders associated with them) are processed correctly within the system.

• Conversations are stored in the database after interactions; this enables retrieval of chat history based on conversation IDs for future reference.

• The final test involves querying both product listings and personal orders successfully, showcasing comprehensive functionality within the application.