​
Definitions

  • Gateway LLM: The gateway LLM is the first LLM that is invoked when the user interacts with your application. Like traditional API Gateways, the gateway LLM is responsible for planning how to process the user’s request and invoke other LLMs, tools or sub-agents to generate the final response.
  • Presentation Layer LLM: The presentation layer LLM is the LLM that is used to generate the UI response. It may or may not be the same as the gateway LLM.

​
C1 as the Gateway LLM (Preferred)

In this pattern, C1 is used as a gateway LLM that can be used to generate responses to user queries which can then call other tools, LLMs or sub-agents to generate the final response. This is the recommended pattern for most applications as it allows you to use the full power of C1 and results in no additional latency over your existing LLM. In a practical example, let’s say you are currently using OpenAI GPT 4.1 as your gateway LLM. Rather than calling the OpenAI API directly, you can use C1 endpoint with a model version that is built on top of GPT 4.1. All your system prompts, tools, etc. would remain the same and the only change needed would be to use the <C1Component> component to render the response.
This is the preferred pattern for integrating Generative UI into your existing applications.

​
C1 as a Presentation Layer LLM

In some cases, you might not want to replace your gateway LLM with C1 due to a variety of reasons (for example if you are using a custom-model). In such cases, you can use the presentation layer LLM pattern where your existing LLM first generates a text-only response and then C1 is used to generate a UI response based on the text response. Pros:
  • Can support any LLM since now the business logic layer LLM is separate from the presentation layer LLM.
Cons:
  • Introduces additional latency since now you’ll have to wait for the entire text response to be generated before you C1 can start streaming
  • Generally speaking, the UI generation works better when C1 has more context of the current state and available functionality. Since the text response is generated by a different LLM, C1 will not have access to these details which it could have used to generate a more accurate UI response.
This pattern is not recommended for most applications as it results in additional latency and is not as flexible as the gateway LLM pattern.

​
C1 as a Tool

In this pattern your gateway LLM decides when to invoke C1 which is exposed to it as a tool call. When the gateway LLM invokes C1, it will pass the current state of the application to C1 as context which it can use to generate a more accurate UI response. This has to be the final tool call in the chain of tool calls and the response returned by C1 should be directly streamed back to the UI without any additional processing. Examples:
  • The user asks to generate a report on the sales of the last 30 days which invokes C1 as a tool call.
  • Your agent needs input from the user to execute its task so it generates a live form via C1 to collect the required information.
Pros:
  • Can support any LLM since now the business logic layer LLM is separate from the presentation layer LLM.
Cons:
  • Introduces additional latency since now you’ll have to wait for the entire text response to be generated before you C1 can start streaming
  • Since the decision to invoke C1 is made by the gateway LLM, it is more error prone and requires more maintenance.