AI-Powered Unsplash to Pinterest Workflow with RAG

Overview

This workflow provides an intelligent pipeline for processing and storing data related to Unsplash images for Pinterest. It uses a Retrieval-Augmented Generation (RAG) architecture to automatically ingest content, generate embeddings, store them in a vector database (Supabase), and then process queries using an AI agent. The workflow is triggered via a webhook, making it suitable for integration with external applications or manual testing.

Node-by-Node Breakdown

1. Sticky Note — A visual note for documentation purposes. It displays the title "Automated workflow: Unsplash to Pinterest" on the canvas. No authentication required.

2. Webhook Trigger — Listens for incoming HTTP POST requests at the path /unsplash-to-pinterest. This is the entry point for the workflow. When triggered, it passes the incoming data to the next nodes. (No auth)

3. Text Splitter — Splits incoming text into chunks of 400 characters with a 40-character overlap. This prepares the data for embedding by breaking it into manageable pieces. (No auth)

4. Embeddings (Cohere) — Generates vector embeddings for each text chunk using the embed-english-v3.0 model from Cohere. These embeddings represent the semantic meaning of the text. (API Key auth — requires a Cohere API key)

5. Supabase Insert — Inserts the generated embeddings into a Supabase vector store named unsplash_to_pinterest. This stores the data for later retrieval. (API Key auth — requires Supabase credentials)

6. Supabase Query — Queries the same Supabase vector store (unsplash_to_pinterest) to retrieve relevant vectors based on similarity search. This is used by the RAG agent to find context. (API Key auth — requires Supabase credentials)

7. Vector Tool — Wraps the Supabase vector store as a tool that the AI agent can use to retrieve context. It's labeled "Vector context" for clarity. (No auth — uses the Supabase connection from the previous nodes)

8. Window Memory — Provides a buffer window memory for the AI agent, allowing it to maintain conversation context across multiple interactions. (No auth)

9. Chat Model (OpenAI) — The language model that powers the AI agent. It uses OpenAI's chat model to process queries and generate responses. (API Key auth — requires an OpenAI API key)

10. RAG Agent — The core AI agent that combines the chat model, vector tool, and memory. It processes incoming data with the prompt: "Process the following data for task 'Unsplash to Pinterest':" and includes a system message: "You are an assistant for Unsplash to Pinterest". (No auth — uses connections to other nodes)

11. Append Sheet (Google Sheets) — Appends the AI agent's output to a Google Sheet. It writes to a sheet named "Log" in a document identified by SHEET_ID. The column "Status" is populated with the agent's response text. (OAuth2 — requires Google Sheets API credentials)

12. Slack Alert — Sends an error notification to the #alerts Slack channel if the RAG Agent encounters an error. The message includes the error details. (OAuth2 — requires Slack API credentials)

Setup Instructions

To use this workflow, you'll need accounts and API keys for the following services:

• Cohere: Sign up at cohere.com and generate an API key for the embeddings model.
• Supabase: Create a project at supabase.com and set up a vector store with the name unsplash_to_pinterest. You'll need your Supabase URL and service role key.
• OpenAI: Get an API key from platform.openai.com for the chat model.
• Google Sheets: Create a Google Sheet with a sheet named "Log" and note the Sheet ID from the URL. Set up OAuth2 credentials in the Google Cloud Console.
• Slack: Create a Slack app, add it to your workspace, and obtain OAuth tokens. Create a channel named #alerts for error notifications.

Configure each node's credentials in n8n by clicking on the node and selecting "Add Credential" or selecting existing ones.

Use Cases and Variations

This workflow is ideal for:

• Content Curation: Automatically process and store descriptions of Unsplash images for later retrieval and Pinterest posting.
• Knowledge Base: Build a searchable database of image metadata with semantic search capabilities.
• Automated Content Moderation: Use the AI agent to analyze image descriptions and flag inappropriate content before posting.

Possible Adaptations:

• Replace the webhook trigger with a Schedule Trigger to run the workflow at regular intervals.
• Swap the Cohere embeddings with OpenAI embeddings or Hugging Face embeddings.
• Instead of Google Sheets, output to Airtable, Notion, or PostgreSQL.
• Add an HTTP Request node after the RAG Agent to post processed data directly to Pinterest's API.
• Use a different vector store like Pinecone or Qdrant instead of Supabase.

{
  "name": "Unsplash to Pinterest",
  "nodes": [
    {
      "parameters": {
        "content": "Automated workflow: Unsplash to Pinterest",
        "height": 530,
        "width": 1100,
        "color": 5
      },
      "id": "8da48de7-ef4d-4d78-8c0e-3c6b57ef3b9a",
      "name": "Sticky Note",
      "type": "n8n-nodes-base.stickyNote",
      "typeVersion": 1,
      "position": [
        -480,
        -240
      ]
    },
    {
      "parameters": {
        "httpMethod": "POST",
        "path": "unsplash-to-pinterest"
      },
      "id": "672c0eab-878f-484d-a072-9482935437c4",
      "name": "Webhook Trigger",
      "type": "n8n-nodes-base.webhook",
      "typeVersion": 1,
      "position": [
        -300,
        0
      ]
    },
    {
      "parameters": {
        "chunkSize": 400,
        "chunkOverlap": 40
      },
      "id": "87553792-b816-44e2-ac2f-64d0f9a5adf5",
      "name": "Text Splitter",
      "type": "@n8n/n8n-nodes-langchain.textSplitterCharacterTextSplitter",
      "typeVersion": 1,
      "position": [
        -130,
        0
      ]
    },
    {
      "parameters": {
        "model": "embed-english-v3.0",
// ... truncated (copy to see full JSON)