n8n 07 — AI Agents & LLM Integration

n8n's AI capabilities are built on LangChain. The relevant node categories are:

# 1. Create OpenAI credential:
Settings → Credentials → New → "OpenAI API"
API Key: sk-proj-your-key-here

# 2. Add "OpenAI Chat Model" node (under AI → Chat Models)
Credential: (select the one you created)
Model: gpt-4o           # or gpt-4o-mini, gpt-3.5-turbo
Temperature: 0.7        # 0=deterministic, 1=creative
Max Tokens: 1000        # limit response length
Timeout: 60000ms

# Simplest AI chat workflow:
When Chat Message Received (trigger)
   └─► AI Agent node
         - Chat Model: OpenAI gpt-4o-mini
         - System Prompt: "You are a helpful assistant."
               └─► (respond automatically via chat interface)

# To expose as a public chat:
# The "When Chat Message Received" trigger gives you
# a chat URL you can embed or share directly.

The AI Agent node can use tools to take real actions. Connect tool nodes to the "Tools" input of the AI Agent.

# AI Agent node settings:
System Message: |
  You are a product support agent.
  Use the search tool to find product info.
  Use the calculator for pricing questions.
  Always cite your sources.

# Connected tools (attach to Tools input):
- Calculator tool    → handles math
- HTTP Request tool  → can call your internal API
- SerpAPI tool       → web search capability
- Code execution     → run small scripts

# Example agent reasoning trace:
User: "What's the price of 5 units of SKU-123 with 10% discount?"
Agent: [calls Calculator] → 5 × $49.99 × 0.9 = $224.95
Agent: "That would be $224.95 for 5 units with 10% discount."

# Phase 1: Ingest documents (run once)
HTTP Request / Google Drive → load PDF
   └─► Default Data Loader (extract text)
         └─► Recursive Character Text Splitter
               (Chunk Size: 1000, Overlap: 100)
               └─► OpenAI Embeddings
                     └─► Pinecone Vector Store (Insert)

# Phase 2: Query (on each user question)
When Chat Message Received
   └─► AI Agent
         - Chat Model: OpenAI gpt-4o
         - System: "Answer only from the provided context."
         - Tool: Vector Store Retriever
                 (connects to Pinecone, returns top 5 chunks)
               └─► Reply with answer + source references

# Postgres Chat Memory setup:
1. Add "Postgres Chat Memory" node
2. Connect it to the "Memory" input of the AI Agent
3. Set Session ID: {{ $sessionId }}  (auto from chat trigger)
4. Set Table Name: n8n_chat_memory
5. Connect your Postgres credential

# The agent now remembers the full conversation history
# across multiple messages in the same session

Email Summarizer

Gmail Trigger (new email arrives)
   └─► OpenAI Chat Model (Basic LLM Chain)
         System: "Summarize this email in 2 sentences.
                  Extract: sender, action needed, urgency (low/med/high)"
         User: {{ $json.text }}
               └─► Set Node (store summary + urgency)
                     └─► IF urgency = high
                           └─► Slack: alert team

Support Ticket Classifier

Chain multiple AI nodes to build multi-step reasoning pipelines:

# Step 1: Extract structured data from raw text
LLM Chain 1:
  Prompt: "Extract name, company, budget from: {{ $json.email_body }}"
  Output: { name: "Alice", company: "Acme", budget: "$5000" }

# Step 2: Score the lead based on extracted data
LLM Chain 2:
  Prompt: "Given this lead: {{ $json }}
           Rate their likelihood to buy (1-10) and explain why."
  Output: { score: 8, reason: "Budget matches, clear intent" }

# Step 3: Draft a personalized outreach email
LLM Chain 3:
  Prompt: "Write a personalized intro email to {{ $json.name }}
           at {{ $json.company }} referencing their interest."