Building generative AI? Get ready for generative UI

There’s currently no stopping generative AI spending. Analyst firm Omdia expects spending on generative AI applications to hit $58 billion by 2028. Yet while this level of investment is a significant amount, actual deployment of generative AI applications is still slow. According to research by Deloitte, 68% of organizations have moved 30% or less of their generative AI pilots into production.

The challenge here for generative AI is that users expect a lot more from these kinds of services than they are currently ready to deliver. While a generative AI application might return a response to a user prompt, how useful is that material to consume? Is it in a format that makes the most out of that information, or does it look the same as a standard web application search for static data?

To make the most of generative AI, we should look at generative user interfaces, or generative UI, as well.

What is generative UI?

Generative UI takes the concept of generative AI and applies it to how we interact with data or systems. Just as generative AI makes data interactive and available in natural language, or creates new images or sound in response to a prompt, so generative UI builds interactive context into how data is displayed, depending on what you are asking for. The goal is to deliver the content that the user wants but also in a format that makes the most of that data for the user too.

For instance, what are you really doing when you search for information? This depends on what you know and how precise your search terms are. Traditional search uses keywords, which are useful when you know the exact target, but do not necessarily fit when users search with natural language terms instead. Do you know exactly what you want to find, or are you looking for general examples or a whole category of things? It’s the difference between searching for a precise brand and product name, like “Nike Air Max 90s,” or for general characteristics, like “red training shoes.” In entertainment, it would be the difference between looking up a specific film title and searching for genre or actor.

When you search with less specific terms, returning a list of items or a chat response might not be useful on its own. Instead, you may want to offer more pictures or other forms of display. By understanding how a user searches, you can then alter how results get displayed and add appropriate formats like images or videos. In the example of searching for a movie or new shoes, you could link to sources of film trailers or other video content, giving the user more information so they can make a better choice.

Delivering generative UI

To deliver generative UI, you will have to link up your application with your generative AI components, like your large language model (LLM) and sources of data, and with the tools you use to build the site like Vercel and Next.js. For generative UI, by using React Server Components, you can change the way that you display the output from your LLM service. These components can deliver information that is updated in real time, or is delivered in different ways depending on what formats are best suited to the responses.

As you create your application, you will have to think about some of the options that you might want to deliver. As a user asks a question, the generative AI system must understand the request, determine the appropriate function to use, then choose the appropriate React Server Component to display the response back.

Typically, applications use static rendering for responses that are created at build time or in the background. This data can be pushed to a content delivery network, then served from that location when it is needed. Alternatively, you can use a dynamic rendering approach where the specific result is created at request time. This is useful when you expect results to be heavily personalized or where the result itself can’t be prepared in advance. This approach can also be used with dynamic APIs, where your application will get data from those APIs as part of building the response.

To create the GenerativeUI element, you can add instructions in your prompt to guide how the AI application handles specific requests, ensuring it uses the appropriate functions so that the data in the response can be displayed in the right form. When you have real-time data as part of that request, you can then use that data in the response back to the user. Depending on how you will use the responses, you can also include output formats like making the results into JSON. This also can make it easier to control the results that you get, making them more deterministic rather than different every time.

To create responses with more relevant data, you will need a database that can act as the source for any real time data that you want to include. Any set of data that you want to use has to be turned into vector embeddings to start with. Using a database like this means that you have a definitive source for any information that you provide to your LLM, rather than letting it create responses that are not relevant or accurate. Using a separate database also means that you are aware of how up-to-date that information is and where the data inside comes from.

For these kinds of applications, creating a workflow and knowing how users will interact with the application or service will be essential to the design. Using services like Langflow makes it easier to connect up all the components involved, from user prompts, LLMs, and data sets through to integrations and outputs that provide the results to the user.

Delivering user results

Generative AI provides users with search responses in natural language. However, it is also possible to use generative AI to create new formats for delivering the results back to users, and to make responses more interactive. By thinking through how users might interact with an application and the kinds of responses they will expect, you can put together more extensive and interactive experiences that evolve to meet user expectations.

Taking this further, one of the big selling points for generative AI is personalization, where the user gets responses that are based on their identity, account history, and preferences. With generative AI, personalization promises to be more than simple pattern matching and recommendations, and to incorporate fully interactive responses. However, to make this work, you must consider the context for how that personalization is delivered. This helps to overcome the challenge where personalization techniques miss the mark and where the user feels that the system does not know them well. It becomes an opportunity for organizations to differentiate themselves from the crowd by leveraging what they know about their end users.

By looking at how a user has interacted with an application or service over time, you can build in personalized results that include reactions to both historical behavior and any “in the moment” inputs that you receive. This includes looking at how the user prefers to get their information delivered and the UI to that information, not just the data.

Building generative UI into your applications can be an effective way to deliver better experiences to users. By iterating on what generative AI does well, generative UI can help you create new ways for users to interact and benefit from your service.

Dom Couldwell is head of field engineering EMEA at DataStax.

—

Generative AI Insights provides a venue for technology leaders—including vendors and other outside contributors—to explore and discuss the challenges and opportunities of generative artificial intelligence. The selection is wide-ranging, from technology deep dives to case studies to expert opinion, but also subjective, based on our judgment of which topics and treatments will best serve InfoWorld’s technically sophisticated audience. InfoWorld does not accept marketing collateral for publication and reserves the right to edit all contributed content. Contact doug_dineley@foundryco.com.