A Comprehensive Guide to WilmerAI Workflow Nodes

A Comprehensive Guide to WilmerAI Workflow Nodes

This document provides an exhaustive reference for all available non-memory-related nodes within the WilmerAI workflow system. It is designed to be a complete guide for an LLM to understand the purpose, configuration, and proper usage of each node.

Core LLM Interaction: The Standard Node

The Standard Node (also referred to as Conversational) is the fundamental building block for all direct interactions with a Large Language Model (LLM). Its primary purpose is to assemble a prompt from various sources of context (conversation history, previous node outputs, static text), send it to a specified LLM backend, and process the response.

Properties

Limitations and Key Usage Notes

• Variable Support: Full variables are supported in the systemPrompt and prompt fields. The endpointName, preset, and maxResponseSizeInTokens fields support LIMITED variables (only {agent#Input} from parent workflows and static variables defined in the workflow JSON, NOT {agent#Output} which doesn't exist yet). For maxResponseSizeInTokens, the variable must resolve to a valid integer string (e.g., "5000").
• Responder Node: Only one node in a workflow can have returnToUser set to true. If no node is designated, the last node in the workflow automatically becomes the responder.
• Prompt Fallback: The node prioritizes the prompt field. If it's empty, it will fall back to using the conversation history as defined by lastMessagesToSendInsteadOfPrompt.

Full Syntax Example

This example shows a non-responder "thinking" node that synthesizes information from previous nodes and the conversation history.

{
  "title": "LLM Thinking Over to User Request",
  "type": "Standard",
  "systemPrompt": "System Information: Today is {todays_date_pretty}. The user is {human_persona_name}.\n\n<your_profile>\n{agent3Output}\n</your_profile>\n\n<user_profile>\n{agent2Output}\n</user_profile>",
  "prompt": "Please consider the most recent twenty messages of your online conversation with {human_persona_name}:\n\n<recent_conversation>\n{chat_user_prompt_last_twenty}\n</recent_conversation>\n\nPlease think carefully about all of this by answering the following questions:\n- A) How long has it been since the last message?\n- B) What did {human_persona_name} mean in their last message to you?\n- C) Carefully consider what the best way to respond might be.",
  "endpointName": "Thinker-Endpoint",
  "preset": "Thinker-Preset",
  "maxResponseSizeInTokens": 8000,
  "addUserTurnTemplate": true,
  "returnToUser": false,
  "addDiscussionIdTimestampsForLLM": true,
  "useRelativeTimestamps": true
}

Variable Substitution for Child Workflows

When calling a child workflow via CustomWorkflow, you can pass dynamic values for endpointName, preset, and maxResponseSizeInTokens using scoped_variables. This allows the parent workflow to control which endpoint, preset, and response size the child uses.

Parent Workflow (calls the child with dynamic config):

{
  "title": "Run Analysis with Custom Settings",
  "type": "CustomWorkflow",
  "workflowName": "General_Workflow_Replaceable_Endpoint",
  "workflowUserFolderOverride": "_common",
  "scoped_variables": [
    "MyDynamicEndpoint",
    "MyDynamicPreset",
    "6000"
  ]
}

Child Workflow (receives the values as {agent#Input}):

{
  "title": "Responding Agent",
  "type": "Standard",
  "endpointName": "{agent1Input}",
  "preset": "{agent2Input}",
  "maxResponseSizeInTokens": "{agent3Input}",
  "systemPrompt": "...",
  "prompt": ""
}

In this example, the child workflow will use endpoint MyDynamicEndpoint, preset MyDynamicPreset, and generate up to 6000 tokens.

Logic & Control Flow: The Conditional Node

The Conditional node is a control flow utility that evaluates a complex logical expression and returns the string "TRUE" or "FALSE". Its output is designed to be used by a ConditionalCustomWorkflow node to make branching decisions. It supports comparisons, logical operators (AND, OR), and parentheses () for grouping.

Properties

Limitations and Key Usage Notes

• Variable Support: The condition property supports all available workflow variables.
• Type Inference: The node intelligently infers types. Values in quotes (e.g., 'complete') are strings. Unquoted TRUE or FALSE are booleans. Other unquoted values are treated as numbers if possible, otherwise as strings.
• IMPORTANT: Be careful when doing comparisons. The node distinguishes between a boolean and a string. The output of a Conditional node is a boolean TRUE or FALSE. If agent3Output comes from a Conditional node and is TRUE, the comparison "{agent3Output} == 'TRUE'" will be FALSE because (boolean) TRUE is not equal to (string) 'TRUE'. The correct comparison is "{agent3Output} == TRUE".

Full Syntax Example

This example demonstrates complex logic with OR and parentheses, checking for multiple valid states or an override condition.

{
  "title": "Check for valid state or admin override",
  "type": "Conditional",
  "condition": "({agent1Output} == 'Admin') OR ({agent2Output} == 'Approved' AND {agent3Output} == TRUE)"
}

In-Workflow Routing: The ConditionalCustomWorkflow Node

The ConditionalCustomWorkflow Node provides powerful branching logic. It dynamically selects and executes a sub-workflow based on the value of a conditional variable (e.g., the output from a Conditional node). It also supports a default content fallback, preventing the need for an extra workflow file for a simple default response.

Properties

Limitations and Key Usage Notes

• IMPORTANT: Only scoped_variables, UseDefaultContentInsteadOfWorkflow, and conditionalKey support workflow variables.
• You cannot set a variable as a workflow name. Configuration keys like workflow names inside conditionalWorkflows and routeOverrides must be static, hardcoded strings.
• Matching Logic: The match for conditionalWorkflows is case-insensitive. However, the match for routeOverrides keys is case-sensitive and expects the key to be Capitalized.

Full Syntax Example

This node routes to a specialized coding workflow based on the language detected in a previous step.

{
  "title": "Route to a Specific Coding Model",
  "type": "ConditionalCustomWorkflow",
  "is_responder": true,
  "conditionalKey": "{agent1Output}",
  "workflowUserFolderOverride": "_common",
  "conditionalWorkflows": {
    "Python": "PythonCodingWorkflow",
    "JavaScript": "JavaScriptCodingWorkflow",
    "Default": "GeneralCodingWorkflow"
  },
  "UseDefaultContentInsteadOfWorkflow": "I'm sorry, I can only assist with Python and JavaScript at the moment.",
  "scoped_variables": [
    "{agent2Output}"
  ],
  "routeOverrides": {
    "Python": {
      "systemPromptOverride": "You are an expert Python programmer. The user's request is: {agent1Input}"
    },
    "JavaScript": {
      "systemPromptOverride": "You are a master JavaScript developer. The user's request is: {agent1Input}"
    }
  }
}

Values sent in as scoped_variables are accessible within the child workflow that is called as agentXInputs. So the first scoped_variable is agent1Input, second scoped_variable is agent2Input, etc.

Modular Logic: The CustomWorkflow Node

The CustomWorkflow Node allows you to execute a separate workflow from within the current one. This is essential for encapsulating reusable logic and breaking down complex processes. Child workflows run in an isolated context; data must be passed in explicitly via scoped_variables.

Properties

Limitations and Key Usage Notes

• Variable Support: Variable substitution is supported within the scoped_variables array and the legacy prompt override fields.
• Static Configuration: The workflowName and workflowUserFolderOverride fields do not support variables and must be static strings.

Full Syntax Example

This example calls a reusable child workflow to perform a search and summarize the findings.

{
  "title": "Custom Wiki Search: Initial Search",
  "type": "CustomWorkflow",
  "workflowName": "Util_Workflow_Wiki_Search_And_Summarize",
  "workflowUserFolderOverride": "_common",
  "is_responder": false,
  "scoped_variables": [
    "{agent1Output}",
    "{agent3Output}"
  ]
}

Values sent in as scoped_variables are accessible within the child workflow that is called as agentXInputs. So the first scoped_variable is agent1Input, second scoped_variable is agent2Input, etc.

Data Manipulation: The ArithmeticProcessor Node

The ArithmeticProcessor node performs a basic mathematical calculation. It takes a string expression containing two numbers and one operator (+, -, *, /), resolves any variables, computes the result, and returns it as a string. If the expression is invalid, it returns "-1".

Properties

Limitations and Key Usage Notes

• Variable Support: The expression property supports all available workflow variables. These variables are expected to resolve to numeric values.

Full Syntax Example

This node calculates a final price by adding a 7% tax to a subtotal from a previous node.

{
  "title": "Calculate Final Price with Tax",
  "type": "ArithmeticProcessor",
  "expression": "{agent1Output} * 1.07",
  "returnToUser": false
}

Data Manipulation: The StringConcatenator Node

The StringConcatenator is a utility for combining multiple strings into one. It takes a list of strings, resolves any variables, and joins them with a specified delimiter. It can also act as a streaming responder.

Properties

Limitations and Key Usage Notes

• Variable Support: Every string within the strings list supports full variable substitution.

Full Syntax Example

This example assembles a multi-line user profile from various data sources.

{
  "title": "Assemble User Profile Block",
  "type": "StringConcatenator",
  "strings": [
    "--- User Profile ---",
    "Name: {user_name}",
    "Key Trait: {agent1Output}",
    "Report Generated: {todays_date_pretty}"
  ],
  "delimiter": "\n",
  "returnToUser": false
}

Data Extraction: The JsonExtractor Node

The JsonExtractor node extracts a specific field from a JSON string. It parses a JSON object (optionally wrapped in markdown code blocks), resolves workflow variables, and returns the value of the specified field as a string. This is useful for parsing structured LLM outputs.

Properties

Limitations and Key Usage Notes

• Variable Support: Both jsonToExtractFrom and fieldToExtract support full variable substitution.
• Markdown Handling: The node automatically strips ```json and ``` code block wrappers before parsing.
• Return Types: Strings are returned as-is; numbers/booleans become strings; nested objects/arrays become JSON strings; null returns an empty string.

Full Syntax Example

This node extracts the "name" field from a JSON object returned by a previous LLM node.

{
  "title": "Extract Character Name from LLM Response",
  "type": "JsonExtractor",
  "jsonToExtractFrom": "{agent1Output}",
  "fieldToExtract": "name"
}

Example Input (from agent1Output):

{"name": "Socg", "file": "Socg.txt"}

Example Output: Socg

Data Extraction: The TagTextExtractor Node

The TagTextExtractor node extracts content from XML/HTML-style tags within a text string. It searches for a specified tag (e.g., <answer>...</answer>) and returns the content between the opening and closing tags. This is useful for parsing structured LLM outputs where the model wraps content in custom tags.

Properties

Limitations and Key Usage Notes

• Variable Support: Both tagToExtractFrom and fieldToExtract support full variable substitution.
• Case Sensitivity: Tag matching is case-sensitive (<Answer> and <answer> are different).
• First Match: If multiple instances of the tag exist, only the first match is extracted.
• Whitespace: Leading/trailing whitespace is stripped from the extracted content.

Full Syntax Example

This node extracts the content within <answer> tags from an LLM response.

{
  "title": "Extract Answer from LLM Response",
  "type": "TagTextExtractor",
  "tagToExtractFrom": "{agent1Output}",
  "fieldToExtract": "answer"
}

Example Input (from agent1Output):

I've analyzed your question. Here is my response:

<answer>
The capital of France is Paris.
</answer>

I hope this helps.

Example Output: The capital of France is Paris.

Utility: The GetCustomFile Node

The GetCustomFile node loads the content of a local text file into the workflow as a string. This allows you to inject large blocks of static text (like instructions or lore) without cluttering the workflow JSON.

Properties

Limitations and Key Usage Notes

• Variable Support: The filepath field supports full variable substitution, including {Discussion_Id} for per-conversation files and {YYYY_MM_DD} for date-based files.
• File Not Found: If the file doesn't exist, the node returns "Custom instruction file did not exist".
• IMPORTANT: Do not set a delimiter or custom delimiter if you want the file to be pulled as it was originally written.

Full Syntax Example

This node loads a character sheet and replaces a simple --- separator with a more decorative one.

{
  "title": "Load Character Sheet",
  "type": "GetCustomFile",
  "filepath": "C:\\Users\\User\\Desktop\\character_sheet.txt",
  "delimiter": "---",
  "customReturnDelimiter": "\n**********\n"
}

Dynamic Filepath Example

This node loads session-specific notes using the conversation's unique identifier.

{
  "title": "Load Session Notes",
  "type": "GetCustomFile",
  "filepath": "/data/sessions/{Discussion_Id}_notes.txt"
}

Utility: The SaveCustomFile Node

The SaveCustomFile node writes string content to a local text file. This is useful for saving data generated during a workflow, such as an LLM's analysis, a conversation summary, or a report.

Properties

Limitations and Key Usage Notes

• Variable Support: Both filepath and content fields support full variable substitution, including {Discussion_Id} for per-conversation files and {YYYY_MM_DD} for date-based files.
• Error Handling: The node returns a status message indicating success or failure (e.g., due to permissions).
• Directory Creation: If parent directories don't exist, the node will attempt to create them.

Full Syntax Example

This node saves a character bio generated by a previous node to a file.

{
  "title": "Save Character Bio to File",
  "type": "SaveCustomFile",
  "filepath": "D:\\WilmerAI\\Characters\\jax_the_pirate.txt",
  "content": "CHARACTER PROFILE\n-----------------\nName: Jax\nBio: {agent1Output}"
}

Dynamic Filepath Example

This node saves a daily report using the date variable.

{
  "title": "Save Daily Report",
  "type": "SaveCustomFile",
  "filepath": "/data/reports/{YYYY_MM_DD}_report.txt",
  "content": "Report for {todays_date_pretty}:\n\n{agent1Output}"
}

Utility: The StaticResponse Node

The StaticResponse node returns a hardcoded string. It's versatile for debugging, providing static instructions, or delivering canned responses without an LLM call. When designated as a responder, it can simulate a streaming response.

Properties

Limitations and Key Usage Notes

• Variable Support: The content field does not support variable substitution; it is treated as a literal string.
• Streaming: If returnToUser is true and the request is for streaming, the content is delivered word-by-word.

Full Syntax Example

This node sends a pre-written message directly to the user as a final, streaming response.

{
  "title": "Return System Status Message",
  "type": "StaticResponse",
  "content": "Affirmative. All systems are operating within nominal parameters. This is a pre-recorded message.",
  "returnToUser": true
}

Vision: The ImageProcessor Node

The ImageProcessor node is the bridge between user-provided images and text-based nodes. It calls a vision-capable LLM to generate text descriptions of images in the user's latest message. These descriptions are then made available to subsequent nodes.

Properties

Limitations and Key Usage Notes

• Limited Variable Support: endpointName and preset support LIMITED variables (only {agent#Input} from parent workflows and static variables, NOT {agent#Output} which doesn't exist yet).
• Sequential Processing: The node processes images one by one and aggregates their descriptions into a single string output, separated by \n-------------\n.
• [IMAGE_BLOCK] Placeholder: This special keyword is only valid inside the message property of this node. It marks where the aggregated descriptions will be inserted.

Full Syntax Example

This node analyzes all user images and injects the description into the chat history for a subsequent text-only node to use.

{
  "title": "Analyze and Describe All User Images",
  "type": "ImageProcessor",
  "endpointName": "Vision-Endpoint",
  "preset": "Vision-Preset",
  "addAsUserMessage": true,
  "message": "[SYSTEM: An image analysis module has processed the user's recent image(s). The detailed description is below:\n\n[IMAGE_BLOCK]\n\nThis is now part of our conversation.]",
  "systemPrompt": "You are a world-class visual analysis AI. Describe the image in meticulous detail for a text-only assistant.",
  "prompt": "Based on our recent conversation:\n{chat_user_prompt_last_five}\n\nDescribe the image's contents in extreme detail. Transcribe any text you see."
}

External Tools: The PythonModule Node

The PythonModule node executes a custom Python script, allowing you to extend workflow capabilities with custom logic, API calls, or file access. The script must contain an Invoke(*args, **kwargs) function that returns a single string value.

Properties

Limitations and Key Usage Notes

• Variable Support: All values within the args array and kwargs object support variable substitution.
• Script Requirement: The target Python file must define a function Invoke(*args, **kwargs) that returns a single value, which will be converted to a string.
• Output Variable: The string returned by the Invoke function becomes the node's output.

Full Syntax Example

This node executes a script to process data, passing arguments from workflow variables.

{
  "title": "My Custom Python Tool",
  "type": "PythonModule",
  "module_path": "C:/WilmerAI/Public/Scripts/process_data.py",
  "args": [
    "A static string argument",
    "{agent1Output}"
  ],
  "kwargs": {
    "api_key": "your-secret-key",
    "user_id": "{userName}"
  }
}

External Tools: The Offline Wikipedia Nodes

The Offline Wikipedia nodes query a local OfflineWikipediaTextApi service to retrieve factual information from a Wikipedia database. This is used for Retrieval-Augmented Generation (RAG) by providing context to an LLM. The promptToSearch field supports variabl

…(truncated)