n8n AI/LLM Cluster Node System

n8n integrates with LangChain to provide advanced AI capabilities via a cluster node architecture — root nodes connected to specialized sub-nodes through typed connectors. Requires n8n v1.19.4+.

Quick Reference

Cluster Node Architecture

AI workflows in n8n use root nodes (agents, chains) connected to sub-nodes (models, memory, tools) through typed AI connectors. Root nodes NEVER work alone — they ALWAYS require at least one Chat Model sub-node.

┌─────────────────────────────────────────────────┐
│  ROOT NODE (Agent/Chain)                        │
│  ┌──────────┬──────────┬──────────┬───────────┐ │
│  │ai_language│ai_memory │ai_tool   │ai_output  │ │
│  │Model     │          │          │Parser     │ │
│  └────┬─────┴────┬─────┴────┬─────┴─────┬─────┘ │
└───────┼──────────┼──────────┼───────────┼───────┘
        │          │          │           │
   ┌────▼────┐ ┌──▼───┐ ┌───▼────┐ ┌───▼──────┐
   │Chat     │ │Memory│ │Tool    │ │Output    │
   │Model    │ │Node  │ │Node(s) │ │Parser    │
   └─────────┘ └──────┘ └────────┘ └──────────┘

AI Node Type Reference

Sub-Node Connection Types (NodeConnectionType)

Decision Trees

Which Agent Type to Use

Need autonomous AI reasoning?
├─ YES: Does the task require tool use?
│  ├─ YES: Which provider?
│  │  ├─ OpenAI with function calling → OpenAI Functions Agent
│  │  ├─ Any provider, general tools → Tools Agent (RECOMMENDED default)
│  │  └─ Need step-by-step planning → Plan and Execute Agent
│  └─ NO: Simple conversation?
│     ├─ YES → Conversational Agent
│     └─ NO: Need reasoning trace? → ReAct Agent
├─ Database queries? → SQL Agent
└─ NO: Simple prompt-response?
   ├─ Single prompt → Basic LLM Chain
   ├─ Summarize text → Summarization Chain
   └─ Q&A over documents → Retrieval QA Chain

Rule: ALWAYS start with Tools Agent unless you have a specific reason to use another type. It is the most flexible and works with any chat model provider.

Which Memory Type to Use

Need conversation memory?
├─ NO → Skip memory sub-node entirely
├─ YES: Persistence required?
│  ├─ NO (in-memory only):
│  │  ├─ Simple buffer → Simple Memory (default 5 exchanges)
│  │  └─ Token-limited → Token Buffer Memory
│  └─ YES (survives restarts):
│     ├─ PostgreSQL available → PostgresChat Memory
│     ├─ Redis available → Redis Chat Memory
│     ├─ Need summarization → Summary Memory
│     └─ Managed service → Zep or Xata Memory

Which Vector Store to Use

Need vector similarity search?
├─ Testing/prototyping → In-Memory Vector Store
├─ Production:
│  ├─ Managed cloud service:
│  │  ├─ Pinecone (fully managed, scalable)
│  │  ├─ Qdrant (open-source, self-hostable)
│  │  ├─ Weaviate (hybrid search)
│  │  └─ Azure AI Search (Azure ecosystem)
│  ├─ Existing database:
│  │  ├─ PostgreSQL → PGVector
│  │  ├─ Supabase → Supabase Vector Store
│  │  ├─ MongoDB → MongoDB Atlas
│  │  └─ Redis → Redis Vector Store
│  └─ Self-hosted → Chroma or Milvus

Core Patterns

Pattern 1: Basic Agent Workflow

[Trigger] → [Tools Agent]
                ├── ai_languageModel → [OpenAI Chat Model]
                ├── ai_memory → [Simple Memory]
                └── ai_tool → [Calculator]
                             [Wikipedia]
                             [Custom Code Tool]

ALWAYS connect at least one Chat Model sub-node. NEVER leave the ai_languageModel connector empty.

Pattern 2: RAG Data Insertion

[Trigger] → [Get Documents] → [Vector Store (Insert Documents)]
                                  ├── ai_embedding → [OpenAI Embeddings]
                                  └── ai_document → [Default Data Loader]
                                                       └── ai_textSplitter → [Recursive Character Text Splitter]

ALWAYS use a text splitter when inserting documents. NEVER insert full documents without splitting — it degrades retrieval quality.

Text splitting guidance:

• ALWAYS use Recursive Character Text Splitter as the default choice
• Use chunk sizes of 200-500 tokens for fine-grained retrieval
• ALWAYS set overlap (10-20% of chunk size) to preserve context across boundaries

Pattern 3: RAG Retrieval via Agent

[Chat Trigger] → [Tools Agent]
                     ├── ai_languageModel → [OpenAI Chat Model]
                     ├── ai_memory → [Postgres Chat Memory]
                     └── ai_tool → [Vector Store Q&A Tool]
                                       └── ai_vectorStore → [Pinecone]
                                                               └── ai_embedding → [OpenAI Embeddings]

Pattern 4: RAG Retrieval via Chain

[Chat Trigger] → [Retrieval QA Chain]
                     ├── ai_languageModel → [OpenAI Chat Model]
                     └── ai_retriever → [Vector Store Retriever]
                                            └── ai_vectorStore → [PGVector]
                                                                    └── ai_embedding → [OpenAI Embeddings]

Pattern 5: Human-in-the-Loop

[Chat Trigger] → [Tools Agent]
                     ├── ai_languageModel → [Chat Model]
                     └── ai_tool → [Tool with Approval]
                                       ├── Approve → [Execute Action]
                                       └── Deny → [Notify User]

• 9 notification channels: Chat, Slack, Discord, Telegram, Microsoft Teams, Gmail, WhatsApp, Google Chat, Microsoft Outlook
• Access tool context: $tool.name (tool identifier), $tool.parameters (AI-determined values)
• Use $fromAI() for dynamic parameter specification in tool nodes
• ALWAYS include human review information in the system prompt so the AI understands the approval workflow

Critical Rules

ALWAYS

• ALWAYS connect a Chat Model sub-node to every agent and chain root node
• ALWAYS use the same embedding model for insertion AND retrieval in RAG workflows
• ALWAYS use Recursive Character Text Splitter unless you have a specific reason not to
• ALWAYS set chunk overlap when splitting documents for RAG
• ALWAYS use Tools Agent as the default agent type
• ALWAYS include a system prompt that describes available tools and expected behavior
• ALWAYS test AI workflows with pinned data before activating in production

NEVER

• NEVER mix embedding models between insertion and retrieval — vectors become incompatible
• NEVER skip text splitting when inserting documents into vector stores
• NEVER connect sub-nodes to incompatible connector types (e.g., a memory node to an ai_tool connector)
• NEVER use Basic LLM Chain when you need tool use — use an Agent instead
• NEVER store sensitive data in AI memory without considering data retention policies
• NEVER use In-Memory Vector Store in production — data is lost on restart
• NEVER assume AI agent output is deterministic — ALWAYS validate critical outputs

Sub-Node Connection Rules

supplyData() Method

AI sub-nodes implement supplyData() instead of execute(). This method returns the LangChain object (model, memory, tool, etc.) that the root node consumes:

// AI sub-node pattern (e.g., a memory node)
async supplyData(this: ISupplyDataFunctions, itemIndex: number): Promise<SupplyData> {
    const memory = new BufferMemory({ /* config */ });
    return { response: memory };
}

Root nodes call getInputConnectionData() to retrieve sub-node outputs:

// Inside agent/chain root node
const model = await this.getInputConnectionData('ai_languageModel', itemIndex);
const memory = await this.getInputConnectionData('ai_memory', itemIndex);
const tools = await this.getInputConnectionData('ai_tool', itemIndex);

LangChain Code Node

The LangChain Code node provides special built-in methods for custom LangChain operations. These methods are ONLY available in the LangChain Code node, NOT in regular Code nodes.

Use the LangChain Code node when:

• Built-in AI nodes do not cover your use case
• You need custom LangChain chain composition
• You need advanced prompt engineering beyond what the UI supports

Reference Links

• AI Node Types and Methods — Complete node catalog with providers and parameters
• AI Workflow Examples — Agent, RAG, and tool usage workflow patterns
• AI Anti-Patterns — Common mistakes and how to avoid them