Prompt Nodes

Build prompts and run generation from graphs: load and render prompt templates, assemble prompts dynamically (see Prompt Building), and generate a response with a client.

7 nodes.

Node	Registry path
Build Prompt	`prompt/BuildPrompt`
Clean Response	`prompt/CleanResponse`
Generate Response	`prompt/GenerateResponse`
Load Template	`prompt/LoadTemplate`
Prompt From Template	`prompt/PromptFromTemplate`
Render Prompt	`prompt/RenderPrompt`
Template Variables	`prompt/TemplateVariables`

Build Prompt

prompt/BuildPrompt

Builds a prompt from a jinja2 template, exposing common context building blocks (scene, memory, extra context, dynamic instructions etc.) as template variables that can be toggled through the node's properties.

Loads the template {scope}.{template_file} and renders it against the active scene using the agent's client for token budgeting. Outputs both the Prompt object and the rendered prompt text.

Inputs

Input	Type	Description
`state`	`any`	The graph state
`agent`	`agent`	The agent whose client is used for rendering and token budgeting
`instructions`	`str`	Task instructions made available to the template (optional)
`dynamic_context`	`list`	List of DynamicInstruction objects injected as extra context (optional)
`dynamic_instructions`	`list`	List of DynamicInstruction objects injected as instructions (optional)
`memory_prompt`	`str`	Semantic query / retrieval prompt for the memory context (optional)

Outputs

Output	Type	Description
`state`	`state`	The state input, passed through
`agent`	`agent`	The agent input, passed through
`prompt`	`prompt`	The built Prompt object
`rendered`	`str`	The rendered prompt text
`response_length`	`int`	The configured response length, passed through

Properties

Property	Type	Default	Description
`template_file`	`str`	`"base"`	The template file to use
`scope`	`str`	`"common"`	The scope of the template
`instructions`	`text`	`""`	The instructions to include in the prompt
`reserved_tokens`	`int`	`312`	The number of tokens to reserve to account for any overhead. Range: 16 – 1024.
`limit_max_tokens`	`int`	`0`	Limit the maximum number of tokens used for the prompt (0 = client context limit). Min: 0.
`include_scene_intent`	`bool`	`True`	Include the scene intent
`include_extra_context`	`bool`	`True`	Include the extra context (pins, reinforcements, content classification)
`include_memory_context`	`bool`	`True`	Include the memory context
`include_scene_context`	`bool`	`True`	Include the scene context
`include_character_context`	`bool`	`False`	Include the active character context
`include_gamestate_context`	`bool`	`False`	Include the game state context
`memory_prompt`	`str`	`""`	Semantic query / retrieval prompt for memory
`prefill_prompt`	`str`	`""`	Prefill the prompt with a response
`return_prefill_prompt`	`bool`	`False`	Return the prefill prompt with the response
`dedupe_enabled`	`bool`	`True`	Enable deduplication
`response_length`	`int`	`0`	The length of the response
`technical`	`bool`	`False`	Include the technical context where applicable (ids, typing etc.)

Clean Response

prompt/CleanResponse

Cleans a response

Inputs

Input	Type	Description
`response`	`str`	The response to clean

Outputs

Output	Type	Description
`cleaned`	`str`	The cleaned response

Properties

Property	Type	Default	Description
`partial_sentences`	`bool`	`True`	partial_sentences
`strip_partial_sentences`	`bool`	`True`	Strip partial sentences from the response

Generate Response

prompt/GenerateResponse

Sends a prompt to the agent and generates a response

Retries up to attempts times on an empty response. When a response_spec is provided its extractors are applied to the response; otherwise, if the template produced a data structure, that is used as the extracted value.

Inputs

Input	Type	Description
`state`	`any`	The graph state
`agent`	`agent`	The agent to send the prompt to
`prompt`	`prompt`	The prompt to send to the agent
`action_type`	`str`	Classification of the generated response (optional)
`response_length`	`int`	The maximum length of the response (optional)
`response_spec`	`response/spec`	(optional) Optional ResponseSpec for extracting structured data from response

Outputs

Output	Type	Description
`state`	`state`	The state input, passed through
`agent`	`agent`	The agent that generated the response
`prompt`	`prompt`	The Prompt object, passed through
`response`	`str`	The response from the agent
`data_obj`	`dict,list`	The data structure of the response (when data_output is enabled)
`captured_context`	`str`	Context captured by the template during rendering
`rendered_prompt`	`str`	The rendered prompt
`response_spec`	`response/spec`	Pass-through of the input response spec
`extracted`	`dict`	Dictionary of extracted values (when response_spec provided)

Properties

Property	Type	Default	Description
`data_output`	`bool`	`False`	Output the response as a data structure
`data_multiple`	`bool`	`False`	Allow multiple data structures in the response
`response_length`	`int`	`256`	The maximum length of the response
`action_type`	`str`	`"scene_direction"`	Classification of the generated response. Choices: `analyze`, `conversation`, `create`, `edit`, `narrate`, `scene_direction`, `summarize`, `visualize`, `world_state`.
`attempts`	`int`	`1`	The number of attempts (retry on empty response)

Load Template

prompt/LoadTemplate

Loads the raw unrendered jinja2 template content based on scope and name

Inputs

Input	Type	Description
`name`	`str`	(optional) The template name (without .jinja2 extension)
`scope`	`str`	(optional) The template scope (optional, defaults to property)

Outputs

Output	Type	Description
`template_content`	`str`	The raw unrendered template content as a string
`scope`	`str`	The resolved template scope, passed through

Properties

Property	Type	Default	Description
`scope`	`str`	`"scene"`	The template scope. Choices are generated at runtime.
`name`	`str`	`""`	The template name to load (without .jinja2 extension)

Prompt From Template

prompt/PromptFromTemplate

Loads a talemate template prompt

Either a template file or raw template text must be provided (but not both).

Inputs

Input	Type	Description
`template_file`	`str`	The template file to load (optional)
`template_text`	`str`	Raw template text to use instead of a file (optional)
`variables`	`dict`	The variables to use in the template (optional)

Outputs

Output	Type	Description
`prompt`	`prompt`	The Prompt instance

Properties

Property	Type	Default	Description
`scope`	`str`	`"scene"`	The template scope. Choices are generated at runtime.
`template_file`	`str`	`""`	The template to load
`template_text`	`text`	`""`	The template text to use
`dedupe`	`bool`	`True`	Enable prompt deduplication

Render Prompt

prompt/RenderPrompt

Renders a prompt

Inputs

Input	Type	Description
`prompt`	`prompt`	The prompt to render

Outputs

Output	Type	Description
`rendered`	`str`	The rendered prompt

Template Variables

prompt/TemplateVariables

Helper node that returns a pre defined set of common template variables

Variables:

scene: The current scene
scene_title: The scene title (falls back to the scene name)
max_tokens: The maximum number of tokens in the response
agent: The relevant agent

Inputs

Input	Type	Description
`agent`	`agent`	The relevant agent
`merge_with`	`dict`	A dictionary of variables to merge with the pre defined variables (optional)

Outputs

Output	Type	Description
`variables`	`dict`	A dictionary of variables
`agent`	`agent`	The relevant agent