n8n-ai-features（n8nAI 特点）

n8n AI Features

Build AI-powered automation with n8n's AI nodes and LangChain integration.

AI Agents

Available Agent Types

Tools Agent

Most flexible - uses tools/functions to complete tasks.

{
  "agent": "toolsAgent",
  "options": {
    "systemMessage": "You are a helpful assistant that can search and analyze data."
  }
}

• Best for: General-purpose tasks, API integrations
• Works with: Any tool node (HTTP, database, code)

Conversational Agent

Memory-enabled for multi-turn conversations.

{
  "agent": "conversationalAgent",
  "options": {
    "systemMessage": "You are a customer support agent.",
    "maxIterations": 10
  }
}

• Best for: Chatbots, support systems
• Works with: Memory nodes for context

ReAct Agent

Reasoning + Acting - thinks step-by-step.

{
  "agent": "reAct",
  "options": {
    "maxIterations": 15,
    "returnIntermediateSteps": true
  }
}

• Best for: Complex reasoning tasks
• Shows chain-of-thought

SQL Agent

Database-focused with schema awareness.

{
  "agent": "sqlAgent",
  "options": {
    "topK": 10,
    "dialect": "postgresql"
  }
}

• Best for: Natural language to SQL
• Works with: Database credentials

Plan and Execute Agent

Plans first, then executes steps.

{
  "agent": "planAndExecute",
  "options": {
    "humanInputMode": "never"
  }
}

• Best for: Multi-step complex tasks
• More deliberate than ReAct

LangChain Nodes

Chains

Basic LLM Chain

Simple prompt → response.

[Chat Trigger] → [Basic LLM Chain] → [Output]
                       ↑
               [OpenAI Chat Model]

Configuration:

{
  "promptType": "define",
  "text": "Summarize this text: {{ $json.content }}"
}

Question and Answer Chain

RAG-based Q&A with context retrieval.

[Input] → [Retriever] → [Q&A Chain] → [Answer]
               ↑              ↑
        [Vector Store]  [LLM Model]

Summarization Chain

Long document summarization.

{
  "type": "map_reduce",
  "options": {
    "chunkSize": 4000
  }
}

Types:

• stuff - Single pass (short docs)
• map_reduce - Chunk and combine (long docs)
• refine - Iterative refinement

Information Extractor

Structured data extraction from text.

{
  "text": "={{ $json.document }}",
  "schema": {
    "type": "object",
    "properties": {
      "company": {"type": "string"},
      "revenue": {"type": "number"},
      "employees": {"type": "integer"}
    }
  }
}

Text Classifier

Categorize text into predefined labels.

{
  "categories": ["positive", "negative", "neutral"],
  "text": "={{ $json.review }}"
}

Sentiment Analysis

Built-in sentiment detection.

{
  "text": "={{ $json.feedback }}",
  "options": {
    "returnScore": true
  }
}

LLM Providers

OpenAI

{
  "model": "gpt-4-turbo",
  "options": {
    "temperature": 0.7,
    "maxTokens": 2000,
    "topP": 1
  }
}

Models: gpt-4-turbo, gpt-4, gpt-3.5-turbo

Anthropic Claude

{
  "model": "claude-3-opus-20240229",
  "options": {
    "temperature": 0.5,
    "maxTokens": 4096
  }
}

Models: claude-3-opus, claude-3-sonnet, claude-3-haiku

Google Gemini

{
  "model": "gemini-pro",
  "options": {
    "temperature": 0.7
  }
}

Ollama (Distributed Cluster)

{
  "baseUrl": "http://100.124.63.99:11434",
  "model": "qwen2.5:32b",
  "options": {
    "temperature": 0.8
  }
}

• Primary: Gaming PC (100.124.63.99) - RTX 4070 GPU
• Fallback: Studio (localhost:11434) - M2 Max
• Always-on: Tower (tower.local:11434) - CPU
• Models: qwen2.5, raz, deepseek-r1, llava, etc.

Groq

{
  "model": "mixtral-8x7b-32768",
  "options": {
    "temperature": 0.5
  }
}

• Ultra-fast inference
• Good for high-volume

Mistral

{
  "model": "mistral-large-latest",
  "options": {
    "temperature": 0.7
  }
}

Vector Stores

Supabase Vector

{
  "tableName": "documents",
  "queryName": "match_documents",
  "options": {
    "matchThreshold": 0.8,
    "matchCount": 5
  }
}

Setup:

-- Enable pgvector extension
create extension if not exists vector;

-- Create documents table
create table documents (
  id uuid primary key default gen_random_uuid(),
  content text,
  metadata jsonb,
  embedding vector(1536)
);

-- Create matching function
create function match_documents(
  query_embedding vector(1536),
  match_threshold float,
  match_count int
)
returns table (id uuid, content text, similarity float)
language sql stable
as $$
  select id, content, 1 - (embedding <=> query_embedding) as similarity
  from documents
  where 1 - (embedding <=> query_embedding) > match_threshold
  order by embedding <=> query_embedding
  limit match_count;
$$;

Pinecone

{
  "indexName": "my-index",
  "namespace": "documents",
  "options": {
    "topK": 5
  }
}

Qdrant

{
  "collectionName": "documents",
  "url": "http://localhost:6333",
  "options": {
    "limit": 10,
    "scoreThreshold": 0.7
  }
}

PGVector

{
  "tableName": "embeddings",
  "options": {
    "distanceStrategy": "cosine",
    "k": 5
  }
}

In-Memory Vector Store

{
  "memoryKey": "document_store"
}

• Good for prototyping
• Lost on restart

Embeddings

OpenAI Embeddings

{
  "model": "text-embedding-3-small",
  "options": {
    "batchSize": 512,
    "stripNewLines": true
  }
}

Models:

• text-embedding-3-small (1536 dims, cheaper)
• text-embedding-3-large (3072 dims, better)
• text-embedding-ada-002 (legacy)

Cohere Embeddings

{
  "model": "embed-english-v3.0",
  "inputType": "search_document"
}

Ollama Embeddings (Gaming PC)

{
  "baseUrl": "http://100.124.63.99:11434",
  "model": "nomic-embed-text"
}

Memory Nodes

Buffer Memory

Simple conversation history.

{
  "sessionKey": "={{ $json.userId }}",
  "contextWindowLength": 10
}

Buffer Window Memory

Limited window of recent messages.

{
  "sessionKey": "chat_{{ $json.sessionId }}",
  "windowSize": 5
}

Motorhead Memory

External memory service.

{
  "url": "http://localhost:8080",
  "sessionId": "={{ $json.userId }}"
}

Zep Memory

Advanced memory with search.

{
  "baseUrl": "http://localhost:8000",
  "sessionId": "user_123"
}

RAG (Retrieval-Augmented Generation)

Basic RAG Pattern

[Document Loader] → [Text Splitter] → [Embeddings] → [Vector Store]
                                                           ↓
[User Query] → [Retriever] → [Context + Query] → [LLM] → [Response]

Document Loading

[HTTP Request] → [Extract Text] → [Text Splitter]
[Read File] → [Extract PDF] → [Text Splitter]
[Google Drive] → [Download] → [Text Splitter]

Text Splitters

Character Text Splitter

{
  "chunkSize": 1000,
  "chunkOverlap": 200,
  "separator": "\n\n"
}

Recursive Character Splitter

{
  "chunkSize": 1000,
  "chunkOverlap": 200,
  "separators": ["\n\n", "\n", ". ", " "]
}

• Better for preserving context
• Tries larger separators first

Token Splitter

{
  "chunkSize": 500,
  "chunkOverlap": 50,
  "encodingName": "cl100k_base"
}

• Token-accurate for LLM context

Retrieval Configuration

{
  "topK": 5,
  "scoreThreshold": 0.7,
  "searchType": "similarity"
}

Search types:

• similarity - Cosine similarity
• mmr - Maximum Marginal Relevance (diversity)

AI Workflow Patterns

Chatbot with Memory

[Chat Trigger]
      ↓
[Buffer Memory] ←→ [Conversational Agent]
                          ↓
                   [OpenAI Model]
                          ↓
                   [Response]

Document Q&A System

[Webhook: /upload]           [Webhook: /query]
      ↓                            ↓
[PDF Extract]                [Supabase Retriever]
      ↓                            ↓
[Text Splitter]              [Q&A Chain]
      ↓                            ↓
[Embeddings]                 [OpenAI Model]
      ↓                            ↓
[Supabase Store]             [Response]

Content Generation Pipeline

[Schedule: Daily]
      ↓
[HTTP: Get Topics] → [For Each Topic]
                          ↓
                   [Basic LLM Chain: Generate Article]
                          ↓
                   [Basic LLM Chain: Edit/Refine]
                          ↓
                   [HTTP: Post to CMS]

AI-Powered Data Processing

[Webhook: Data Input]
      ↓
[Information Extractor]
      ↓
[Code: Validate/Transform]
      ↓
[If: Confidence > 0.8]
      ↓
[Database: Insert]

Multi-Model Routing

[Input]
   ↓
[Text Classifier: Complexity]
   ↓
[Switch]
   ├→ simple → [GPT-3.5]
   ├→ medium → [GPT-4-Turbo]
   └→ complex → [Claude Opus]
         ↓
     [Merge]
         ↓
     [Output]

MCP (Model Context Protocol) Integration

MCP Client Node

Connect to MCP servers for extended capabilities.

{
  "serverUrl": "http://localhost:3000/mcp",
  "tools": ["web_search", "calculator", "code_interpreter"]
}

Best Practices

1. Prompt Engineering

// Use structured prompts
const systemPrompt = `You are a helpful assistant.

RULES:
1. Be concise
2. Use bullet points
3. Cite sources

FORMAT:
- Summary: [brief answer]
- Details: [expanded explanation]
- Sources: [references]`;

2. Temperature Settings

3. Token Management

• Monitor token usage per execution
• Use summarization for long contexts
• Implement chunking for large documents
• Cache embeddings to reduce API calls

4. Error Handling

[AI Node]
   ↓ (on error)
[Error Trigger]
   ↓
[Fallback Response]
   ↓
[Log Error]

5. Cost Optimization

• Use smaller models for simple tasks
• Cache frequent queries
• Batch similar requests
• Use local models (Ollama) for development

Output Format

AI WORKFLOW:
  Purpose: [What it automates]
  LLM: [Model choice and why]
  Vector Store: [If RAG, which store]

NODES:
  1. [Trigger] - [Configuration]
  2. [AI Node] - [Model, temperature, prompt]
  ...

PROMPT TEMPLATE:
  System: [System message]
  User: [User message template]

CONSIDERATIONS:
  - Tokens: [Estimated usage]
  - Cost: [Per execution estimate]
  - Latency: [Expected response time]