Skip to main content

​
Introduction

The Nadoo AI Workflow Engine is a LangGraph-based visual workflow system that lets you build AI agent pipelines as directed graphs. Each workflow is a graph where nodes process inputs, retrieve information, make decisions, and generate responses, connected by edges that define the execution flow. Whether you are building a simple chatbot or a multi-step RAG pipeline with tool use, the workflow engine provides the runtime, streaming, and observability you need to go from prototype to production.

​
Node Categories

Nadoo AI ships with 18+ built-in node types organized into five categories.

Input / Output

Nodes that receive user input and deliver final responses.
  • Start Node — Entry point for every workflow
  • End Node — Terminal node that finalizes output
  • Question Node — Prompt the user for additional information
  • Form Node — Collect structured data via form fields
  • Direct Reply Node — Return an immediate static or templated response

AI / LLM

Nodes that invoke language models and multimodal AI.
  • AI Agent Node — Core LLM interaction with 6 execution modes
  • Image Generate — Create images from text prompts
  • Image Understand — Analyze and describe images
  • TTS (Text-to-Speech) — Convert text to audio
  • STT (Speech-to-Text) — Transcribe audio to text

Knowledge & Retrieval

Nodes for RAG, search, and data access.
  • Search Knowledge Node — Vector / hybrid search over knowledge bases
  • Reranker Node — Re-score and reorder retrieved documents
  • Document Extract Node — Parse and extract content from files
  • Database Node — Execute SQL queries against relational databases
  • Database Semantic RAG Node — Natural-language-to-SQL with retrieval

Logic & Control

Nodes for branching, variables, and sub-workflows.
  • Condition Node — If/else branching based on expressions
  • Variable Node — Set, transform, or compute variables
  • Application Node — Invoke another Nadoo application as a sub-workflow

Integration

Nodes for external tools, code, and MCP servers.
  • Tool Node — Call a registered tool by name
  • Tool Lib Node — Access shared tool libraries
  • Python Node — Execute arbitrary Python code
  • MCP Node — Connect to Model Context Protocol servers

​
Execution Model

Every workflow execution is governed by two layers of context:
ContextScopePurpose
WorkflowContextGlobal (entire run)Holds the conversation history, global variables, streaming channel, and execution metadata
NodeContextLocal (single node)Contains node-specific inputs, configuration, and intermediate results
Data flows between nodes through the global WorkflowContext. Each node reads its inputs from the context, performs its operation, and writes its outputs back for downstream nodes to consume.

​
Node Lifecycle

Every node passes through a consistent lifecycle during execution:
1

Pre-execute

Validate inputs, resolve variable references, and prepare the node’s runtime configuration.
2

Execute

Run the node’s core logic — call an LLM, query a knowledge base, evaluate a condition, etc.
3

Post-execute

Process outputs, update the workflow context with results, and record metrics.
4

Resolution

Determine which downstream edge(s) to follow based on the node’s output and any branching conditions.

​
Node Status States

During execution, each node transitions through the following statuses:
StatusDescription
PENDINGNode is queued and waiting to execute
RUNNINGNode is currently executing
SUCCESSNode completed successfully
FAILEDNode encountered an error
SKIPPEDNode was bypassed (e.g., condition evaluated to false)
INTERRUPTEDNode was stopped by user intervention or timeout

​
SSE Streaming

The workflow engine streams execution events to the client in real time via Server-Sent Events (SSE). This enables live progress indicators, token-by-token LLM output, and detailed debugging. There are 19 event types covering the full execution lifecycle:
  • workflow_started — Execution begins
  • workflow_finished — Execution completed successfully
  • workflow_failed — Execution terminated with an error
  • workflow_interrupted — Execution stopped by user

​
Example: Minimal Workflow

A simple question-answering workflow requires only three nodes:
  1. Start Node receives the user’s message.
  2. AI Agent Node sends the message to an LLM and streams the response.
  3. End Node delivers the final answer.

​
Next Steps

Visual Editor

Learn how to build workflows with the drag-and-drop editor

AI Agent Node

Deep dive into the most important node type

AI Agent Strategies

Explore the 6 execution modes for LLM reasoning

Knowledge Base

Set up RAG pipelines for your workflows