createLLMStreamNode

workflow.createLLMStreamNode() Creates a streaming LLM invocation node. When the node is executed, provider.stream(llmInput) is called, LLMStreamChunk is produced, and runStreamNode collects chunks and returns aggregate text.

Signature

workflow.createLLMStreamNode<Input>(
  options: Pick<
    LLMNodeOptions<Input>,
    "name" | "provider" | "mapInput" | "historyLimit" | "metadata"
  >,
): BaseStreamNode<Input, LLMStreamChunk, string>

Parameter

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

Options Field

Field	Type	Description
name	string	Node name, default "llm-stream".
provider	LLMProviderSource	provider instance, provider id, provider ref, or resolver. Required.
mapInput	(input, context) => LLMChatInput	Translate the business input into an LLM request. Required.
historyLimit	number	The number of nodes retained in the execution history. The default value is 50.
metadata	WorkflowNodeMetadataInput	The node displays meta information.

LLMStreamChunk

Type	Field	Description
text-delta	text: string	A piece of incremental text.
finish	finishReason?: FinishReason	Stream end reason.
finish	usage?: LanguageModelUsage	token usage.
finish	providerMetadata?: ProviderMetadata	provider meta information.

Return value

Return	Type	Description
Node Object	BaseStreamNode<Input, LLMStreamChunk, string>	Execute via workflow.runStreamNode.
chunks	LLMStreamChunk[]	runStreamNode All chunks collected.
output	string	Node finalize splices all text-delta.text.

When executor is configured with tokenUsage: { enabled: true }, the token usage is automatically reported from the usage of the stream finish chunk by using the provider parsed by getLLMProvider() during node execution. the same stream is reported only once. if the provider does not return the usage, the token usage is not written. Automatic reporting retains the input, output, reasoning, cached input, and cache creation/write token details.

Call process

Phase	Input	Result
Upstream node	Input	Business input
mapInput	Input, NodeContext	LLMChatInput
provider stream	LLMChatInput	AsyncIterable<LLMStreamChunk>
runStreamNode	chunks	{ chunks, output: string }

Example

const streamNode = workflow.createLLMStreamNode<ParsedInput>({
  name: "draft-stream",
  provider: workflow.createLLMProviderRef("default"),
  mapInput(input) {
    return {
      prompt: input.message,
      temperature: 0.3,
    };
  },
});

const { chunks, output } = await workflow.runStreamNode(streamNode, parsed, context);

Difference from createLLMNode

Method	Output method	Suitable scenarios
createLLMNode	One-time return of LLMChatOutput or mapOutput results	Background processing, non-real-time output.
createLLMStreamNode	Chunk-by-chunk output, finally aggregated into a string	Streaming text needs to be shown in Desktop/server/embed.