createLLMNode

workflow.createLLMNode() Creates a one-time LLM invocation node. When the node executes, it will first parse the provider, then call provider.chat(llmInput), and finally convert the LLM output into business payload through mapOutput.

Signature

workflow.createLLMNode<Input>(
  options: LLMNodeOptions<Input, LLMChatOutput>,
): BaseNode<Input, LLMChatOutput>

workflow.createLLMNode<Input, Output>(
  options: LLMNodeOptions<Input, Output> & {
    mapOutput: LLMOutputMapper<Input, Output>;
  },
): BaseNode<Input, Output>

Parameter

Parameter	Type	Description
options	LLMNodeOptions<Input, Output>	Define provider, input mappings, output mappings, and node meta information.

LLMNodeOptions

Field	Type	Description
name	string	Node name, default "llm". provider binding to override the provider by that name.
provider	LLMProviderSource	provider instance, provider id, provider ref, or resolver. Required.
mapInput	(input, context) => LLMChatInput	Translate the business input into an LLM request. Required.
mapOutput	(output, input, context) => Output	Converts the LLM output to business output; returns the original LLMChatOutput when not passed.
historyLimit	number	The number of nodes retained in the execution history. The default value is 50.
metadata	NodeMetadataInput	The node displays meta information.

LLMProviderSource

Form	Type	Description
provider examples	LLMProvider	Use the callable provider directly.
provider id	string	Lookup from context.providers.llmProviders.
provider ref	LLMProviderRef	Delay parsing by id and configure the fallback.
resolver	(context, nodeName) => LLMProvider | undefined	Dynamically selects provider by node or context.

LLMChatInput

Field	Type	Description
prompt	string	Single prompt; and messages two-choice 1.
messages	ModelMessage[]	Multi-round message; 1 with prompt binary.
system	string	System prompt.
temperature	number	Sampling temperature.
maxOutputTokens	number	The maximum number of output tokens.
topP / topK	number	Sampling parameters.
presencePenalty / frequencyPenalty	number	Repeat the penalty parameter.
stopSequences	string[]	Stop the sequence.
seed	number	Random seed.
maxRetries	number	Number of retries.
timeout	number	The call timed out.
headers	object	Extra request header.
providerOptions	LLMProviderOptions	provider private extension parameters.
abortSignal	AbortSignal	Nodes are automatically injected from context.abortSignal.

Output

Situation	Node Output Type	Description
Not passed mapOutput	LLMChatOutput	Returns the provider chat result as it is.
Imported mapOutput	Output	Returns the result of mapOutput(output, input, context).

LLMChatOutput

Field	Type	Description
text	string	Model text output.
finishReason	FinishReason | undefined	Stop the cause.
usage	LanguageModelUsage | undefined	token usage.
providerMetadata	ProviderMetadata | undefined	provider meta information.

When an executor configures tokenUsage: { enabled: true }, the provider resolved by getLLMProvider() reports LLMChatOutput.usage to context.tokenUsage after chat returns, so accounting still works when mapOutput omits usage. The run aggregate appears in report.tokenUsage.run; cumulative totals are written to the current project's KV and, when present, its conversation KV. Server derives the system total by aggregating project totals. Input, output, reasoning, cached-input, and cache-creation token details are supported.

Call process

Phase	Input	Result
Upstream node	Input	Business input
mapInput	Input, NodeContext	LLMChatInput
provider chat	LLMChatInput	LLMChatOutput
mapOutput	LLMChatOutput, Input, NodeContext	Output

Example

const llmNode = workflow.createLLMNode<ParsedInput, AnswerPayload>({
  name: "answer",
  provider: workflow.createLLMProviderRef("default"),
  mapInput(input) {
    return {
      system: "You answer briefly.",
      prompt: input.message,
      temperature: 0.2,
    };
  },
  mapOutput(output) {
    return {
      errCode: 0,
      errMessage: "",
      content: output.text,
    };
  },
});