December 19, 2023

Introduction

Threekit's first initiative in integrating AI into a Threekit configurator took place in April 2023 with our shoe configurator. At that time, AI was generating a lot of hype, and it was crucial for us to determine if there were practical use cases for Threekit and the emerging AI tools in the market.

While the project was ultimately successful, we gained valuable insights:

Online shopping differs from in-person experiences

Traditional product configurators may obscure product complexity but often fail to inspire buyers. Online shopping can involve endless scrolling through catalogs, keyword guessing games, or clicking through configurator forms until you finally get to see what you're looking for.

Recommending the "right" product variant is subjective to the buyer

Every individual has unique preferences. While certain products may gain popularity or fall out of fashion, not everyone wants the most popular product configuration.

Powerful conversational configurators must utilize real product data

Threekit's initial exploration of AI, as illustrated by the shoe example, demonstrated that Threekit can work with LLMs (Language Model Models) to create conversational configurators. However, the shoe example was relatively simple. Based on user input, our GPT utility would convert a string like:

User: "I want a shoe that's designed for a fan of the Chicago Bears"

/*
    GPT then queries for data about the Chicago Bears & their colors.
    Per our instructions, it will return a hex code we can apply to our shoe
*/

let colors = ["#0A162A", "#C73703"]

// Set the Threekit config with GPT's outputs
configurator.setConfiguration({'Body Mesh': '#C73703', 'Tongue Color': '0A162A'})

Phase II

While the small POC was great, there are several considerations to provide real value in a conversational configurator:

• The AI concierge should respond with values that can be used by APIs.
• These values should correspond to real Threekit/Product attributes rather than arbitrary responses like hex codes.
• There should be a mechanism to ensure the suggested configurations are valid.

Threekit's AI Initiative

Threekit aims to use AI in our experiences to get buyers as close to 'add-to-cart' as quickly as possible.

When a buyer visits a store and is greeted by a brand representative, stylist, or concierge, they don't always say, "I want this product in blue and red. I like those colors." The conversation is more casual, with descriptions of themselves or the person they're shopping for, their styles, likes, and interests. They rely on the representative to translate their requirements into a product recommendation.

How can we replicate this experience in a digital environment? In our previous examples, we used off-the-shelf OpenAI APIs for testing purposes, but these concepts should apply to other LLMs.

The Data

Threekit Configurable Product Data

Threekit operates as a RAG (retrieval-augmented generation) framework for retrieving relevant information about a configurable product. This framework grounds the model in the realm of product data used in a visual configurator.

We utilize the available product information, including thumbnails, swatches, and additional media, to provide the AI model with detailed information and descriptions for each product.

Training Instructions

Once the AI concierge is provided with the product data and descriptors, we give it instructions to guide its personality and define its goals. This is where you can shape the energy and tone of the conversations with buyers.

What kind of attitude would you like your concierge to have? Is the goal to have the buyer add the product to their cart, or is there a different call-to-action? All of these details can be decided in the instructions.

LLM & The Conversation

The conversation flow is dependent on the instructions & desires of the client but this is one path to take.

• A buyer sends a message.
• If there is enough information in the message to make a recommendation, the concierge will do so.
• If there is not enough information, the concierge will ask the buyer some questions to gain insight.
• Once a decision has been made, the concierge will respond with a combination of attributes and their values, which will be reflected in the Threekit player.
• From here, the buyer can be prompted with the desired call-to-action or further configure the product through a configuration form.

In Conclusion

The journey of integrating AI into Threekit's configurator, starting with the shoe, has been a valuable learning experience. It has illuminated the potential and challenges of infusing AI into our existing systems, shedding light on the distinctions between online and in-person shopping experiences and the subjectivity inherent in product variant recommendations. The successful implementation of GPT in our configurator has showcased the feasibility of using Language Model Models (LLMs) to create dynamic, conversational interfaces that can adapt to individual user preferences and input.

As we look ahead, our primary focus remains on refining the AI configurator to enhance its effectiveness and user-friendliness. This entails ensuring that the AI's responses align with API requirements and that the recommended configurations are practical. This project has underscored the significance of utilizing real product data to inform the AI model, enabling it to offer pertinent and precise recommendations.

Internally, this endeavor serves as an instructive case study on the integration of AI in product configuration. It imparts valuable insights into the process of assimilating new technologies to improve our tools and services, the importance of comprehending user needs in online environments, and the potential of AI to reshape how we interact with and present our products to customers. As we continue to advance this technology, it remains essential to draw lessons from our experiences and refine our approach to meet the evolving demands of online retail and product customization.