AI Discovery
Enable users to search for product recommendations using plain language queries.
Overview
The AI Discovery feature unlocks a powerful way for buyers to interact with the product configuration experience. Instead of relying on a step-by-step UI that walks users through a set of predefined choices, in order to find their desired product configuration, the AI Discovery process enables users to get a set of recommendations based on a plain language text input. Simply making a request such as "I am looking for a three seat sofa with a durable fabric for my cat, in a fresh spring color." could be sufficient to get a set of recommendations, similar to the experience of an in-person chat with a sales person.
AI Discovery is fully dependent on the Catalog 2.0 Items, Options, and Attributes. Classic Catalog Items and Attributes will NOT function with the experience.
The platform currently supports only one AI Discovery Experience which, once created, cannot be deleted.
A proper AI Agent setup is critical for a successful user experience.
AI Agent Setup
The AI Discovery Experience is powered by a set of AI agents who need to process the user query and then explain to the user why the results are relevant. These AI agents require guidance in order to be effective in their jobs.
Requirements
The AI agent can only provide meaningful options to the end user if it has a good understanding of the catalog of products. This understanding will come from two main sources - descriptions and variants.
Descriptions
It is critical that all items, attributes, and options in the Catalog 2.0 have their descriptions set with as much context information as possible. While the AI agent can also use the title and label of the items, attributes and options, those will most likely be insufficient to correctly determine what they represent.
Fill out the Description of every Item, Attribute, and Option in the Catalog 2.0 with sufficient context for the AI agent to parse the catalog effectively and come up with appropriate recommendations.
If you need to make dual use of the Descriptions for both the front end UI as well as for AI context, then you may feel like you would be forced to artificially limit your description to only what would be suitable to display on the front-end UI. In this scenario, it would be better to use the Description field for AI context, and for the front end make use of a Data Table instead, which contains the front-end descriptions for each item.
Variants
The AI agent will only parse the item variants for the necessary data. The variants is where all the descriptions will eventually make their way to provide the AI agent with the necessary context. If the variants end up containing conflicting or unclear information, then the AI agent will have difficulty making good recommendations.
The Vector Data for the variants also needs to be generated, in order for the AI to parse the Variants effectively.
All items in the catalog will need to have their variants and Vector Data generated using the variant manager.
The item variants need to encompass the whole set of available configurations.
Since the variant manager is limited to 3000 variants per item, the AI agent will have trouble making good recommendations for products with more than 3000 possible configurations.
Please see the Variant Manager page for more details.
Rephrase Prompt
The rephrase prompt is a set of instructions to the AI agent processing the user query, on how to interpret the user query. This essentially ends up being the text sent to the AI along with the user query.
Example:
You are a product search assistant helping buyers refine their queries to get the most accurate and relevant results from a vector store search.
Your task is to rephrase the buyer's query while following these rules:
Maintain the original intent of the buyer's query.
Remove any vague or ambiguous terms.
Focus on including specific product features, attributes, or use cases when relevant.
Avoid adding unnecessary words or altering the meaning of the query.
The output should be concise, clear, and optimized for search relevance.
If the buyer's query lacks enough detail, make a reasonable assumption to improve specificity.
Confidence Prompt
The confidence prompt is a set of instructions for the AI on how to explain to the user why the provided results are relevant and how they address the user's query.
Example:
Explain to the buyer why the suggested tractor best suits their needs. Try to make them feel like this will help them work more efficiently and effectively. Give the suggestion like you're an NBA announcer and keep it under 300 words.
The AI will then present a set of recommendations with explanations based on the Confidence Prompt instructions.
Moderation
The Enable Moderation toggle will add an additional filter that checks for the use of profanity in the user query. It will not prevent the query from being processed. Instead, it will just flag the query that it was detected as using profanity, and let the front-end team decide how to handle it in the UI.
If the query passed the moderation filter without any profanity being detected, the Query Experience API response will set the passesModeration parameter to true.
Questions
The Questions are not a required component of the AI Discovery experience, but they can help guide both the buyer and the AI agent towards more specific recommendations.
It would be helpful to think of them as a set of questions that a sales person might ask the buyer, in order to refine the search.
Each question requires a Name, along with a set of predetermined choices.
These will be visible on the Experience Preview page, and could be exposed on the front-end of a custom implementation.
Choices
These choices end up acting as additional query text that gets sent to the AI agent that processes the user query. In fact, they simply end up being concatenated to the text input from the user, resulting in a longer string of text that contains the user input first, immediately followed by the choice text.
For example: if the user inputs "Red sofa", and they pick a Choice labelled "Loveseat" to a question about the style of seating, then the resulting query sent to the AI agent will look like this:
Red sofa Loveseat
If the Choice has the Additional Vector Input filled out, then the query will replace the label of the Choice with the Additional Vector Input text instead. If the Additional Vector Input for the Loveseat choice said "The Product should be a Loveseat Sofa", then the resulting query sent to the AI agent will read:
Red sofa The Product should be a Loveseat Sofa
Each choice requires a label at a minimum, which would show up in the Experience Preview UI, but it would be much more helpful if the Additional Vector Input was populated as well, to provide better instructions to the AI agent for how to treat this choice in terms of filtering.
Usage
Once you have set up the Ai Discovery experience, there are two ways to preview and interact with it:
The built-in default Preview Experience page provided by Threekit
Use the API to custom build your own webpage UI to provide users with the query text box, submit button and filtering questions, along with a way to display the query results.
With both of these methods we provide an additional set of helpful UI elements to customize the look and feel of the experience.
UI Elements
The UI Elements are available for customization under the AI Agent Setup tab, as shown below.
Experience Title
The page title that appears on the Preview page in the top left corner, right above the input query text box.
Search Box Placeholder
The placeholder text that appears in the input query textbox on the Preview page.
Custom Logo
Upload your own logo image to have it show up in the top left corner of the Preview page.
Main Color
This would be the color used by the buttons on the Preview page.
Accent Color
This would be the color used by the View Product links on the Preview page.
Preview Experience Page
The Preview Experience page is a bare-bones front-end implementation of the AI Discovery experience. It is accessible from the Preview Experience button in the top right corner of the AI Discovery page.
The Preview Experience page is somewhat similar in purpose to the Item or Asset Preview page, where you have a basic player embed along with a basic configurator form. It's only intended as a preview and testing ground for your AI Discovery Experience settings, so you can test different queries and see what the AI Agent comes up with in terms of recommendations.
It is not intended to act as your actual client-facing page, or necessary in any way other than for testing purposes prior to a custom front-end implementation.
Front-end Implementation
The recommended way to implement the AI Discovery feature is to build your own custom client-facing web page that hosts the experience.
This custom front-end can be built using the AI Discovery API endpoints. All you need is the following:
Get the Experience settings from the platform, to get all the necessary parameters to populate the UI, such as the title, UI Elements, and Questions along with thumbnail links.
Send the user Query to the Experience to be processed, using the Query Experience API.
Process the return from the Query Experience API endpoint and format it to your desire on your client-facing page. The return includes a set of results along with the
confidenceMessagethat describes the reason why these would be great choices.Read carefully through the AI Discovery API page to understand how to use the filter option on the Query Experience API.
Hard Filters
There will be cases where you may wish to force the AI agent to perform its search only within a clearly defined set of categories, instead of relying on the agent itself to determine if those categories are clearly stated in the variant descriptions.
For this scenario you can make use of the filters parameter on the Query Experience API. It contains a set of instructions on how to perform an AI query that is limited to only search through variants that pass the given filters.
These hard filters will not be passed in the query to the AI agent, the way that Question options are passed. Instead, it will force the AI agent to only search for recommendations through a subset of variants that meet the given criteria. Without the hard filters the AI agent will always perform its search through the entire catalog.
Limitations
Products with more than 3000 potential variants will not work very well with AI Discovery. The AI agent can only search the generated variants for information, and individual items can only have up to 3000 variants generated. The use of hard filters, as described above, may help in these situations.
Descriptions are currently limited to 2500 characters. If you need to also use the Descriptions to fill out product and option descriptions on the front-end UI, then put those descriptions in a Data Table instead, and keep the Description field for the AI context only.
Last updated
Was this helpful?