Visual Workflow Engine

One of the core features of NiteMoon AI Platform is the visual workflow engine, which allows users to build complex AI workflows through a drag-and-drop interface, enabling multi-step AI pipelines without writing code.

Key Features

• 13 built-in node types: Covering LLM chat, knowledge retrieval, conditional branching, tool invocation, and more
• Visual orchestration: Drag-and-drop interface for intuitive complex flow construction
• Context passing: Data flow between nodes through ref_inputs configuration
• Token statistics: Real-time tracking of token consumption across the entire workflow
• SSE streaming response: Support for Server-Sent Events streaming output

Node Types

Flow Control Nodes

Node	Type	Description
START	Entry Node	Workflow starting point, configurable welcome message
END	Terminal Node	Workflow endpoint
IF_ELSE	Conditional Branch	Execute different branches based on conditions

LLM Nodes

Node	Type	Description
LLM	Text Chat	Basic text LLM conversation node
MULTIMODAL_LLM	Multimodal	Supports image, audio, and video input

Knowledge Retrieval Nodes

Node	Type	Description
KNOWLEDGE_RETRIEVAL	RAG Retrieval	Retrieve knowledge from vector store
KG_RETRIEVAL	Graph Retrieval	Retrieve knowledge from knowledge graph

Tool Nodes

Node	Type	Description
DOC_EXTRACTOR	Document Extraction	Extract document content
DATETIME_TOOL	Date & Time	Get current date and time
WEB_SEARCH_TOOL	Web Search	Execute web search
HTTP_REQUEST_TOOL	HTTP Request	Send HTTP request
CUSTOM_TOOL	Custom Tool	User-defined tool
COMMAND_EXEC_TOOL	Command Execution	Execute system command

Workflow Execution Flow

graph TD
    A[START] --> B[LLM Node]
    B --> C{IF_ELSE Condition}
    C -->|Condition 1| D[KNOWLEDGE_RETRIEVAL]
    C -->|Condition 2| E[WEB_SEARCH_TOOL]
    D --> F[END]
    E --> F

1. Execution starts from the START node
2. Nodes are executed sequentially following the connection order
3. IF_ELSE nodes select branches based on conditions
4. The output of each node serves as input for the next node
5. Execution completes upon reaching the END node

Technical Implementation

Graph Traversal Algorithm

The workflow engine uses graph traversal algorithms to execute nodes, supporting:

• Sequential execution
• Conditional branching
• Parallel execution (planned)

Context Passing

Data flows between nodes through ref_inputs configuration:

{
  "ref_inputs": {
    "user_query": "{{START.output}}",
    "context": "{{KNOWLEDGE_RETRIEVAL.results}}"
  }
}

Token Statistics

Real-time tracking of token consumption across the entire workflow, including:

• Input token count
• Output token count
• Total token count

Use Cases

Intelligent Customer Service

START -> LLM(Intent Recognition) -> IF_ELSE -> KNOWLEDGE_RETRIEVAL -> LLM(Generate Response) -> END

Data Analysis

START -> DOC_EXTRACTOR -> LLM(Data Extraction) -> HTTP_REQUEST(Data Query) -> LLM(Analysis Report) -> END

Content Generation

START -> WEB_SEARCH_TOOL -> LLM(Content Generation) -> END

Code Examples

Create a Workflow

// Create workflow definition
WorkflowDefinition workflow = new WorkflowDefinition();
workflow.setName("Intelligent Customer Service Workflow");

// Add nodes
WorkflowNode startNode = new WorkflowNode("START", "Start");
WorkflowNode llmNode = new WorkflowNode("LLM", "Intent Recognition");
WorkflowNode endNode = new WorkflowNode("END", "End");

// Connect nodes
workflow.addEdge(startNode, llmNode);
workflow.addEdge(llmNode, endNode);

// Save workflow
workflowService.save(workflow);

Execute a Workflow

// Execute workflow
WorkflowExecution execution = workflowService.execute(
    workflowId,
    Map.of("userQuery", "How to configure the RAG knowledge base?")
);

// Get execution result
String result = execution.getResult();

Related Links

• Multi-Model LLM Integration
• RAG Knowledge Base
• Quick Start