API Reference: LLMNode

The LLMNode is the "brain" of your agent. It is a resilient, state-aware node that calls a generative AI model (like Google's Gemini) to perform a reasoning task.

It automatically formats your prompt with data from the GraphState and writes the model's text output back to the state.

Key Features

• Resilient: Automatically retries on 429 (rate-limit) errors with exponential backoff.

• Cost Auditing: Automatically logs prompt_tokens, output_tokens, and total_tokens to the run_metadata in the history log.

• Stateful: Uses Python's format() string method to dynamically populate your prompt with any value from the GraphState.

Example Usage

# The LLMNode reads `task` from the state
planner_node = LLMNode(
    model_name="gemini/gemini-2.5-pro",
    prompt_template="You are a planner. Your task is: {task}",
    output_key="plan", # Saves the result to state["plan"]
    next_node=...
)

__init__ Parameters

Parameter	Type	Required	Description
model_name	str	Yes	The name of the model (e.g., "gemini-2.5-pro").
prompt_template	str	Yes	An f-string compatible template. Keys in {braces} will be filled from the GraphState.
output_key	str	Yes	The key to save the LLM's text response to in the GraphState (e.g., "draft_answer").
next_node	BaseNode	Yes	The next node to run after this one successfully completes.
max_retries	int	No	The number of times to retry on a 429 error. Default: 3.