Core LLM Interaction and Advanced Templating

A Feature Guide to WilmerAI for Workflow Authors

This document provides a high-level overview of the core features and capabilities available within the WilmerAI system. It is intended to serve as a comprehensive guide for an LLM tasked with authoring workflows, bridging the gap between the project overview and the detailed technical references for nodes and variables.

For each feature, this guide will explain its purpose, its primary use case, the key workflow nodes associated with it, and reference any available in-depth documentation for further details.

Core LLM Interaction and Advanced Templating

The most fundamental capability of WilmerAI is its ability to interact with Large Language Models. This is primarily achieved through the Standard node, which assembles context, sends a request to a configured LLM endpoint, and processes the response.

A key aspect of this feature is the powerful templating engine that allows for the creation of dynamic, context-aware prompts. While simple variable substitution ({my_variable}) is supported by default, you can enable a full Jinja2 templating engine on a per-node basis by adding "jinja2": true. This unlocks advanced logic like loops and conditionals directly within your prompt strings, which is especially useful for formatting the {messages} variable ( the entire conversation history).

• Key Node: Standard
• Detailed Documentation:
  • A Comprehensive Guide to WilmerAI Workflow Nodes
  • Developer Guide: Workflow Node Jinja2 Support

Modular and Reusable Workflows (Nesting)

WilmerAI allows you to build complex processes by breaking them down into smaller, reusable components. This is achieved by "nesting" workflows—running one workflow file as a single step inside another. A Parent workflow uses a * CustomWorkflow node* to call a self-contained Child workflow.

This architecture promotes reusability (e.g., creating a single summarize.json workflow and calling it from anywhere), simplifies complex logic by breaking it into manageable parts, and enables the creation of high-level "orchestrator" workflows. Data is passed from the parent to the child explicitly via the scoped_variables property, which the child receives as {agent#Input} variables.

• Key Node: CustomWorkflow
• Detailed Documentation: Feature Guide: Custom Nested Workflows

Conditional Logic and In-Workflow Routing

Workflows are not limited to a rigid, linear sequence of steps. WilmerAI supports dynamic, non-linear execution paths through conditional logic. This allows a workflow to function as an intelligent agent that can make decisions and route a task to the most appropriate tool based on the current context.

This is primarily accomplished using the ConditionalCustomWorkflow node, which acts as an "if/then" switch. It evaluates a variable (often the output of a previous step) and executes a specific child workflow based on that value. For more complex evaluations, the Conditional node can be used first to evaluate a logical expression (with AND/ OR operators) and return a simple "TRUE" or "FALSE" to drive the routing decision.

• Key Nodes: ConditionalCustomWorkflow, Conditional
• Detailed Documentation:
  • Feature Guide: WilmerAI's In-Workflow Routing
  • In-Workflow Routing: The ConditionalCustomWorkflow Node

Stateful Conversation Memory

WilmerAI features a sophisticated, three-part memory system to provide long-term, stateful context for conversations. This system is designed for performance by separating the slow process of writing memories from the fast process of reading them.

The three components are:

1. Long-Term Memory File: Chronological, summarized chunks of the conversation.
2. Rolling Chat Summary: A single, continuously updated high-level summary of the entire discussion.
3. Searchable Vector Memory: A vector database containing structured memory objects (title, summary, entities) for efficient semantic search (RAG).

Memory "writer" nodes like QualityMemory run in the background to create and update memories, while fast "reader" nodes like VectorMemorySearch or GetCurrentSummaryFromFile retrieve context to be used in a prompt.

• Key Nodes: QualityMemory (writer), VectorMemorySearch (reader), GetCurrentSummaryFromFile (reader), RecentMemorySummarizerTool (reader)
• Detailed Documentation:
  • Feature Guide: WilmerAI's Memory System
  • WilmerAI Workflow Memory Node Catalog

Automated Conversation Timestamps

To provide LLMs with temporal awareness, WilmerAI can automatically inject timestamps into the conversation history. When enabled on a Standard node, this system prepends a timestamp to each message's content before sending it to the LLM.

The feature can be configured to use absolute timestamps (e.g., (Saturday, 2025-09-20 16:30:05)) or relative ones (e.g., [Sent 5 minutes ago]). Additionally, the system provides the {time_context_summary} variable, which gives a high-level natural language summary of the conversation's timeline (e.g., "This conversation started 2 hours ago...").

• Key Node: Standard (with addDiscussionIdTimestampsForLLM flag)
• Detailed Documentation: A Technical Guide to Conversation Timestamps

External Tool Integration

The workflow engine can be extended with custom tools and integrations to external services. This allows you to add capabilities that are not native to the WilmerAI system.

Two primary methods for this are:

1. Custom Python Scripts: The PythonModule node allows you to execute an arbitrary local Python script and use its string output as the result of the node. This is the most flexible way to add custom logic or connect to other APIs.
2. Offline Wikipedia Integration: WilmerAI has a built-in integration with the OfflineWikipediaTextApi service. A family of OfflineWikiApi... nodes allows a workflow to perform a semantic search against a local Wikipedia database to retrieve factual articles for Retrieval-Augmented Generation (RAG).

• Key Nodes: PythonModule, OfflineWikiApiBestFullArticle, OfflineWikiApiTopNFullArticles, etc.
• Detailed Documentation:
  • Feature Guide: WilmerAI's Offline Wikipedia Integration
  • A Comprehensive Guide to WilmerAI Workflow Nodes

Vision Capabilities

WilmerAI can process and understand images provided in a user's message. The ImageProcessor node takes any images from the user's latest turn, sends them to a configured vision-capable LLM, and generates a detailed text description.

The aggregated text description is then made available as the node's output ({agent#Output}). This allows a subsequent, text-only Standard node to use the description as context, effectively giving the entire workflow "sight."

• Key Node: ImageProcessor
• Detailed Documentation: A Comprehensive Guide to WilmerAI Workflow Nodes

In-Workflow Data and File Manipulation

For common data processing and file system tasks, WilmerAI includes several utility nodes. These nodes allow workflows to read and write local files and perform basic data manipulation without needing an LLM call or an external Python script.

• File I/O: The GetCustomFile node reads the contents of a text file, while the SaveCustomFile node writes string content to a file. Both nodes support variable substitution in their filepath fields, including {Discussion_Id} and {YYYY_MM_DD} for per-conversation or date-based file paths.

• Data Processing: The StringConcatenator node joins a list of strings with a specified delimiter, and the ArithmeticProcessor node evaluates a simple mathematical expression.

• Data Extraction: The JsonExtractor node extracts a specific field from a JSON string (automatically handling markdown code block wrappers), and the TagTextExtractor node extracts content from XML/HTML-style tags within text.

• Key Nodes: GetCustomFile, SaveCustomFile, StringConcatenator, ArithmeticProcessor, JsonExtractor, TagTextExtractor

• Detailed Documentation: A Comprehensive Guide to WilmerAI Workflow Nodes

Concurrency Control

For long-running, asynchronous tasks that should not be run simultaneously (like regenerating a conversation summary), WilmerAI provides a locking mechanism. The WorkflowLock node can be used to acquire a named lock. If another workflow execution reaches a node with the same lock ID, it will terminate immediately, thus preventing race conditions and redundant processing. The lock is automatically released after 10 minutes or when the workflow that acquired it completes.

• Key Node: WorkflowLock
• Detailed Documentation: A Comprehensive Guide to WilmerAI Workflow Nodes

Debugging and Performance Monitoring

WilmerAI provides built-in logging to help debug workflows and monitor performance. At the end of each workflow execution, an INFO-level summary is logged showing every node that executed, including timing information.

Example Output:

=== Workflow Node Execution Summary: MainWorkflow ===
Node 1: Standard || 'Prepare Response' || Responder-Endpoint || http://127.0.0.1:5001 || 182.4 seconds
Node 2: GetCustomFile || 'Load Context' || N/A || N/A || 0.1 seconds
Node 3: CustomWorkflow || 'Route to Helper -> HelperWorkflow' || N/A || N/A || 45.2 seconds
=== End of Summary: MainWorkflow ===

Information Displayed:

• Node index: 1-based position in the workflow
• Node type: The type of node (Standard, GetCustomFile, CustomWorkflow, etc.)
• Node name: From the title field, falling back to agentName, or "N/A"
• Endpoint details: The endpoint name and URL for LLM-calling nodes
• Execution time: Time spent executing this node in seconds

For CustomWorkflow and ConditionalCustomWorkflow nodes, the summary also shows the target workflow name(s) to help trace execution through nested workflows.

• Log Level: INFO (visible in standard logging output)
• Use Cases: Identifying slow nodes, debugging workflow execution order, performance optimization