n8n Master Workflows

A vector store enables Retrieval-Augmented Generation (RAG) by matching natural language queries to relevant text chunks. This workflow uses PostgreSQL with PGVector support for production-ready storage.

The RAG Pipeline:

1. Text Splitting: Use the Recursive Character Text Splitter to break large documents into manageable segments (e.g., chunkSize: 2000).
2. Embedding: Use an embedding model (like Embeddings OpenAI) to transform text segments into numerical vectors.
3. Storage: Use the vectorStorePGVector node in insert mode to store these vectors and their metadata in PostgreSQL.
4. Retrieval: When a user asks a question, the system embeds the query and finds the most relevant vectors in the database.

// Example configuration for the Recursive Character Text Splitter
{
  "name": "Recursive Character Text Splitter",
  "type": "@n8n/n8n-nodes-langchain.textSplitterRecursiveCharacterTextSplitter",
  "parameters": {
    "options": {},
    "chunkSize": 2000
  }
}

The AI Agent acts as the central intelligence, providing conversational support by accessing the knowledge base and maintaining conversation context.

Key Components for the Agent:

• Vector Store Tool: Connect the agent to the Postgres PGVector Store using a toolVectorStore node. This allows the agent to perform RAG to answer specific policy questions.
• Chat Memory: Use Postgres Chat Memory to track the sessionId (typically the Telegram chat.id). This allows the agent to remember previous parts of the conversation.
• Language Model: Connect an OpenAI Chat Model to drive the reasoning and response generation.
• System Prompt: Configure the systemMessage (e.g., "You are a helpful assistant for HR and employee policies") to define the agent's persona and boundaries.

// Example AI Agent configuration
{
  "name": "AI Agent",
  "type": "@n8n/n8n-nodes-langchain.agent",
  "parameters": {
    "text": "={{ $json.text }}",
    "options": {
      "systemMessage": "You are a helpful assistant for HR and employee policies"
    },
    "promptType": "define"
  }
}

You can enable an AI Agent to request human assistance by using the Tool Workflow node. This allows the agent to call a sub-workflow when it encounters a query it cannot answer or lacks confidence in.

Implementation Pattern

1. Main Workflow: Create an AI Agent with a Chat Trigger. Attach a Tool Workflow node to the agent's ai_tool input.
2. Tool Configuration: In the Tool Workflow node, set the description to instruct the agent when to use it (e.g., "Use this tool if you don't know the answer...").
3. Sub-workflow: Create a separate workflow starting with an Execute Workflow Trigger. This sub-workflow handles the logic of what happens when the tool is called (e.g., checking for user data or sending a notification to Slack).
4. Data Passing: Use the fields parameter in the Tool Workflow node to pass relevant data from the main workflow (like chatInput) to the sub-workflow.

{
  "name": "dont_know_tool",
  "type": "@n8n/n8n-nodes-langchain.toolWorkflow",
  "parameters": {
    "workflowId": "={{ $workflow.id }}",
    "description": "Use this tool if you don't know the answer to the user's question, or if you're not very confident about your answer.",
    "fields": {
      "values": [
        {
          "name": "chatInput",
          "stringValue": "={{ $('Chat Trigger').item.json.chatInput }}"
        }
      ]
    }
  }
}

This workflow processes incoming Telegram messages by distinguishing between text and audio inputs to ensure a consistent experience for the AI agent.

Message Routing Logic:

1. Trigger: The Telegram Trigger listens for incoming messages.
2. Type Verification: A Switch node checks the keys in $json.message:
   • Text: If the message contains text, it is routed to an Edit Fields node to normalize the format.
   • Voice: If the message contains voice, the workflow uses the Telegram node to fetch the file, then uses the OpenAI node with the audio/transcribe operation to convert the voice to text.
   • Fallback: If the message type is unsupported, a Telegram node sends a fallback response: "I'm not able to process this message type."
3. Unified Input: Both text and transcribed audio are sent to the AI Agent in a standardized format.

// Example OpenAI transcription configuration
{
  "name": "OpenAI",
  "type": "@n8n/n8n-nodes-langchain.openAi",
  "parameters": {
    "resource": "audio",
    "operation": "transcribe",
    "binaryPropertyName": "=data"
  }
}