Node Types Reference

The WDF schema defines 12 node types. 11 are CLI-supported (validate, push, run). The twelfth — document_extraction — is documented below for schema completeness but is not deployable via the CLI.

Each type has a specific purpose, execution mode, and config schema. Several execution nodes also accept optional saveToMemory / memoryFilePath fields (additive run-memory copy). Those fields are forwarded at push time even when not listed on the strict Pydantic config model.

Input Nodes

plain_txt_input

Pauses the workflow and asks the user for free-form text.

Execution mode: INPUT
When to use: The user provides a question, description, or any unstructured text.

user_input:
  type: plain_txt_input
  execution_mode: INPUT
  label: Ask a Question
  config:
    placeholder: Type your question here...

Config fields:

Field	Type	Required	Description
`placeholder`	string	No	Hint text shown in the input field

structured_input

Pauses the workflow and presents a form with typed fields.

Execution mode: INPUT
When to use: You need specific, structured data from the user (names, selections, numbers).

form_input:
  type: structured_input
  execution_mode: INPUT
  label: Customer Details
  config:
    schema:
      type: object
      properties:
        name:
          title: Full Name
          type: string
        priority:
          title: Priority Level
          type: string
          enum: [low, medium, high]
      required: [name]

Config fields:

Field	Type	Required	Description
`schema`	object	Yes	JSON Schema defining the form fields. Use `properties` for fields, `required` for mandatory ones, `enum` for dropdowns, `title` for display labels.

file_upload

Pauses the workflow and asks the user to upload one or more files.

Execution mode: INPUT
When to use: The workflow processes documents, images, or data files.

upload_doc:
  type: file_upload
  execution_mode: INPUT
  label: Upload Document
  config:
    acceptedFormats: [pdf, docx, txt, csv]
    maxFileSize: 10485760

Config fields:

Field	Type	Required	Description
`acceptedFormats`	list[string]	Yes	Allowed file extensions (e.g., pdf, docx, csv)
`maxFileSize`	integer	Yes	Maximum file size in bytes (10485760 = 10 MB)
`saveToMemory`	boolean	No	When `true`, uploaded files are stored in the RLM sandbox memory bucket (`RLM_SANDBOX_MEMORY_BUCKET_NAME`) instead of the default workflow inputs bucket. Defaults to `false`.

Processing Nodes

llm_call

Sends a prompt to an LLM and returns the generated text.

Execution mode: MESSAGES
When to use: You need AI-generated text — summaries, analysis, classification, reports.

summarize:
  type: llm_call
  execution_mode: MESSAGES
  label: Summarize Content
  config:
    model: us.anthropic.claude-sonnet-4-20250514-v1:0
    temperature: 0.3
    maxTokens: 2048
    system_prompt: You are a helpful assistant.
    template: |
      Summarize the following content:
      {{user_input.output.text}}

      Provide a concise summary in 3-5 bullet points.

Config fields:

Field	Type	Required	Description
`model`	string	Yes	LLM model identifier. Use `us.anthropic.claude-sonnet-4-20250514-v1:0` for Claude Sonnet.
`template`	string	Yes	The prompt template. Use `{{slug.output.field}}` for variable references.
`system_prompt`	string	No	System-level instructions for the LLM.
`temperature`	float	No	0.0 (deterministic) to 2.0 (creative). Default 0.7. Use 0.0-0.3 for extraction/classification, 0.3-0.5 for reports, 0.5-0.7 for creative content.
`maxTokens`	integer	No	Maximum tokens in the response. 1024 for short outputs, 2048-4096 for reports, 8192 for long-form.
`saveToMemory`	boolean	No	When `true`, on top of the normal output, also write a JSON copy of it to a file in the run's memory scope. Additive — unlike `api_consumption`'s `saveToMemory`, the normal output is unaffected. Defaults to `false`.
`memoryFilePath`	string	No	Templated relative path under the run memory scope for the additive copy (e.g. `analysis/{{node_id}}.json`). Defaults to `{node_id}/output.json` when omitted. Only used when `saveToMemory` is `true`.

agent

Delegates processing to a registered platform agent. Agents have their own system prompts, tools, and capabilities.

Execution mode: MESSAGES
When to use: You want a pre-configured AI agent to handle the task flexibly, potentially using tools.

support_agent:
  type: agent
  execution_mode: MESSAGES
  label: Customer Support Agent
  config:
    agent_name: Customer Support Agent
    primaryInput: "{{user_input.output.text}}"

Config fields:

Field	Type	Required	Description
`agent_name`	string	Yes*	Name of the platform agent to invoke. Resolved to UUID during push.
`primaryInput`	string	No	Variable reference for the input to send to the agent.
`model`	string	No	Override the agent's default model.
`temperature`	float	No	Override the agent's default temperature.
`maxTokens`	integer	No	Override the agent's default max tokens.
`system_prompt`	string	No	Override the agent's default system prompt.
`use_rlm`	boolean	No	When `true`, route execution through the RLM (beta) sandbox runner instead of the standard agent loop.
`web_tools_enabled`	boolean	No	When `true`, attach `web_search` / `web_fetch` tools to the agent for this node.
`saveToMemory`	boolean	No	When `true`, on top of the normal output, also write a JSON copy of it to a file in the run's memory scope. Additive — the normal output is unaffected. Defaults to `false`.
`memoryFilePath`	string	No	Templated relative path under the run memory scope for the additive copy. Defaults to `{node_id}/output.json` when omitted. Only used when `saveToMemory` is `true`.

*Either agent_name or agentId (UUID) is required.

rag_agent

A retrieval-augmented agent that searches knowledge bases before responding.

Execution mode: MESSAGES
When to use: You need an agent that draws on your documents/knowledge to answer questions.

qa_agent:
  type: rag_agent
  execution_mode: MESSAGES
  label: Knowledge Base Agent
  config:
    agent_name: Research Agent
    knowledge_base_names:
      - Company Policies
      - Industry Standards
    primaryInput: "{{user_input.output.text}}"

Config fields:

Field	Type	Required	Description
`agent_name`	string	Yes*	Name of the platform agent.
`knowledge_base_names`	list[string]	Yes*	Knowledge base names to search. Resolved to UUIDs during push.
`primaryInput`	string	No	Variable reference for input routing.
`topK`	integer	No	Number of documents to retrieve per KB.
`system_prompt`	string	No	Override the agent's system prompt.
`saveToMemory`	boolean	No	When `true`, on top of the normal output, also write a JSON copy of it to a file in the run's memory scope. Additive — the normal output is unaffected. Defaults to `false`.
`memoryFilePath`	string	No	Templated relative path under the run memory scope for the additive copy. Defaults to `{node_id}/output.json` when omitted. Only used when `saveToMemory` is `true`.

retrieve

Performs vector/semantic search against a knowledge base. Returns matching documents.

Execution mode: FLOW
When to use: You need to fetch relevant documents from a knowledge base before processing them with an LLM.

search_kb:
  type: retrieve
  execution_mode: FLOW
  label: Search Knowledge Base
  config:
    knowledge_base_name: Company Policies
    topK: 5
    scoreThreshold: 0.5

Config fields:

Field	Type	Required	Description
`knowledge_base_name`	string	Yes*	KB name. Resolved to UUID during push.
`topK`	integer	No	Number of results to return (default 5).
`searchQuery`	string	No	Custom search query. Can use variable references.
`scoreThreshold`	float	No	Minimum relevance score (0.0-1.0).
`enableReranking`	boolean	No	Enable result reranking (default false).
`includeMetadata`	boolean	No	Include document metadata (default true).
`saveToMemory`	boolean	No	When `true`, on top of the normal output, also write a JSON copy of it to a file in the run's memory scope. Additive — the normal output is unaffected. Defaults to `false`.
`memoryFilePath`	string	No	Templated relative path under the run memory scope for the additive copy. Defaults to `{node_id}/output.json` when omitted. Only used when `saveToMemory` is `true`.

structured_output

Sends a prompt to an LLM and validates the response against a JSON Schema.

Execution mode: OUTPUT
When to use: You need the AI to produce specific structured data (JSON with defined fields).

extract_fields:
  type: structured_output
  execution_mode: OUTPUT
  label: Extract Invoice Fields
  config:
    model: us.anthropic.claude-sonnet-4-20250514-v1:0
    schema:
      type: object
      properties:
        vendor_name:
          type: string
        total_amount:
          type: number
        line_items:
          type: array
          items:
            type: string
      required: [vendor_name, total_amount]
    primaryInput: "{{upload.output.text}}"
    system_prompt: Extract the requested fields from the document.

Config fields:

Field	Type	Required	Description
`schema`	object	Yes	JSON Schema defining the expected output structure.
`model`	string	No	LLM model identifier.
`primaryInput`	string	No	Variable reference for input data.
`system_prompt`	string	No	Instructions for the extraction.
`saveToMemory`	boolean	No	When `true`, on top of the normal output, also write a JSON copy of it to a file in the run's memory scope. Additive — the normal output is unaffected. Defaults to `false`.
`memoryFilePath`	string	No	Templated relative path under the run memory scope for the additive copy. Defaults to `{node_id}/output.json` when omitted. Only used when `saveToMemory` is `true`.

document_extraction

CLI unsupported. Valid in the WDF schema and recognized by workflow validate step 2, but check 10 (Unsupported Node Types) fails. Author these nodes in the Builder UI or via the API.

Extracts structured fields from documents using a field list and optional extraction settings.

Execution mode: MESSAGES
When to use: Legacy structured extraction pipelines (prefer structured_output or llm_call for new workflows).

extract_invoice:
  type: document_extraction
  execution_mode: MESSAGES
  label: Extract Invoice Fields
  config:
    fields:
      - name: vendor_name
        type: string
        required: true
      - name: total_amount
        type: number
        required: true
    extractionMethod: llm
    prompt: Extract the listed fields from the uploaded document.
    extractTables: false
    extractImages: false

Config fields:

Field	Type	Required	Description
`fields`	list[object]	No	Fields to extract. Each entry: `name`, `type`, `required` (boolean).
`extractionMethod`	string	No	Extraction backend hint.
`prompt`	string	No	Instructions for the extraction step.
`extractTables`	boolean	No	Include table extraction when supported.
`extractImages`	boolean	No	Include image extraction when supported.

Human Interaction Nodes

human_review

Pauses the workflow for a human to approve, reject, or request revision.

Execution mode: FLOW
When to use: A person needs to check the AI's work before proceeding.

manager_review:
  type: human_review
  execution_mode: FLOW
  label: Manager Approval
  config:
    review_prompt: >
      Review the generated report. Approve if accurate,
      reject if fundamentally wrong.
    timeoutMinutes: 1440
    allowApprove: true
    allowReject: true
    allowEdit: false

Config fields:

Field	Type	Required	Description
`review_prompt`	string	No	Instructions shown to the reviewer.
`timeoutMinutes`	integer	No	How long to wait before timing out (1440 = 24 hours).
`allowApprove`	boolean	No	Enable approve action (default true).
`allowReject`	boolean	No	Enable reject action (default true).
`allowEdit`	boolean	No	Enable inline editing before approval (default false).

Integration Nodes

api_consumption

Calls an external HTTP API through a configured org-scoped API Connector. The connector (referenced by connectorId) carries the OpenAPI schema, variable definitions, host allowlist, and secrets on the backend.

Execution mode: MESSAGES
When to use: The workflow needs live data from a third-party API, or needs to download a large response body (a transcript, export, or media file) into the run's memory for downstream nodes.

fetch_transcript:
  type: api_consumption
  execution_mode: MESSAGES
  label: Download Zoom Transcript
  config:
    connectorId: zoom-api
    primaryInput: "{{zoom_trigger.output.text}}"
    operationHint: getMeetingTranscript
    timeoutSeconds: 60
    saveToMemory: true
    memoryFilePath: "transcripts/{{zoom_trigger.output.meeting_uuid}}.vtt"

Config fields:

Field	Type	Required	Description
`connectorId`	string	Yes	UUID of the org-scoped API Connector to invoke.
`primaryInput`	string	No	Variable reference for the input routed to the connector.
`maxRecursionDepth`	integer	No	Max follow-up API calls the node may chain (default 1).
`operationHint`	string	No	Name of the connector operation to prefer.
`timeoutSeconds`	integer	No	Per-request timeout in seconds.
`saveToMemory`	boolean	No	When `true`, stream the HTTP response body to a file in the run's memory scope instead of parsing it inline. Defaults to `false`.
`memoryFilePath`	string	No	Templated, path-confined relative path under the run memory scope (e.g. `transcripts/{{trigger.output.meeting_uuid}}.vtt`). Only used when `saveToMemory` is `true`. Defaults to `api/{node_id}/response.<ext>` when omitted. Must be relative — absolute paths and `..` segments are rejected.
`responseVariableMappings`	list of objects	No	Extract JSON paths from the response body and expose them as named output variables. Each entry is `{ variable: <name>, jsonPath: <glom path> }`. `jsonPath` is a glom path into the parsed JSON response (e.g. `types[0].type.name`, `data.results[0].id`, `name`); `variable` is the output name, referenced downstream as `{{node.output.<variable>}}`. Defaults to an empty list.
`headers`	map of string to string	No	Templated per-request HTTP headers (e.g. `authorization: 'Bearer {{get_token.output.access_token}}'`). Merged by the executor with the connector's auth headers, where connector auth wins on collision.
`callParams`	map of string to string	No	Templated query/call parameters (e.g. a `from`/`to` date window) forwarded to the planner.

When saveToMemory is true, the node exposes the response as a memory file rather than inline text. Feed the resulting path into a downstream memory_file_url node (see below) to produce a signed download URL.

See 03-variable-references.md for the output.memory_file_path, output.memory_file_url, output.content_type, output.size_bytes, and output.status_code paths exposed by this node.

Use responseVariableMappings to pull specific values out of a JSON response and expose them as named output variables, instead of parsing the whole body downstream:

fetch_pokemon:
  type: api_consumption
  execution_mode: MESSAGES
  label: Fetch Pokemon
  config:
    connectorId: pokeapi
    primaryInput: "{{user_q.output.text}}"
    operationHint: getPokemon
    responseVariableMappings:
      - variable: primary_type
        jsonPath: "types[0].type.name"
      - variable: base_experience
        jsonPath: "base_experience"

Each mapped variable is then available downstream as {{fetch_pokemon.output.primary_type}} and {{fetch_pokemon.output.base_experience}}. The jsonPath uses glom syntax: dotted keys (data.results) and bracketed list indices (types[0]).

memory_file_url

Produces a signed download URL for a file in the run's memory scope (RLM sandbox memory bucket).

Execution mode: OUTPUT
When to use: End users or downstream nodes need a clickable link for a file written by api_consumption (saveToMemory), file_upload (saveToMemory), or an execution node with additive saveToMemory.

transcript_url:
  type: memory_file_url
  execution_mode: OUTPUT
  label: Transcript Download Link
  config:
    path: "{{fetch_transcript.output.memory_file_path}}"

Config fields:

Field	Type	Required	Description
`path`	string	Yes	Relative path under the org memory root (e.g. `outputs/report.pdf`). May use `{{slug.output.field}}` templates. Must be relative — absolute paths and `..` segments are rejected.

Output paths: output.url, output.filename. See 03-variable-references.md.

Navigation Menu

Uh oh!

Node Types Reference

Input Nodes

plain_txt_input

structured_input

file_upload

Processing Nodes

llm_call

agent

rag_agent

retrieve

structured_output

document_extraction

Human Interaction Nodes

human_review

Integration Nodes

api_consumption

memory_file_url

Uh oh!

FilesExpand file tree

02-node-types-reference.md

Latest commit

History

02-node-types-reference.md

File metadata and controls

Node Types Reference

Input Nodes

plain_txt_input

structured_input

file_upload

Processing Nodes

llm_call

agent

rag_agent

retrieve

structured_output

document_extraction

Human Interaction Nodes

human_review

Integration Nodes

api_consumption

memory_file_url