Stop Coding Prompts: Building a "Self-Correcting" AI Agent with n8n & Supabase

We need a database that understands the "semantic meaning" of a requirement, not just keyword matching.

1. Create a Project at Supabase.
2. Navigate to the SQL Editor and run the following setup script.

This enables the pgvector extension and creates a similarity search function.

-- 1. Enable Vector Extension
create extension if not exists vector;

-- 2. Create the Storage Table
create table examples (
  id uuid primary key,
  content text,       -- The "User Requirement" (what we search against)
  metadata jsonb,     -- The full context (Recommendation + Explanation)
  embedding vector(1024) -- MUST match your embedding model dimensions
);

-- 3. Create the Similarity Search Function
create or replace function relevant_examples (
  query_embedding vector(1024),
  match_threshold float,
  match_count int
)
returns table (
  id uuid,
  content text,
  metadata jsonb,
  similarity float
)
language plpgsql
as $$
begin
  return query
  select
    examples.id,
    examples.content,
    examples.metadata,
    1 - (examples.embedding <=> query_embedding) as similarity
  from examples
  where 1 - (examples.embedding <=> query_embedding) > match_threshold
  order by examples.embedding <=> query_embedding
  limit match_count;
end;
$$;

Instead of managing vectors directly, we use n8n's internal Data Tables as our CMS (Content Management System).

1. In n8n, create a Data Table named Website Recommendation Examples.
2. Add three string columns:
   • requirements (The input scenario)
   • recommendation (The correct output)
   • explanation (The "Chain of Thought")

Seed Data: Add 10-15 distinct examples.

• Row 1: "Simple portfolio for a photographer" → "Squarespace"
• Row 2: "Multi-vendor marketplace like Airbnb" → "Bubble or Sharetribe"

We need a workflow that watches the n8n table and updates the Supabase vector store. This ensures that when you add a new "lesson," the AI learns it automatically.

1. Trigger: Schedule (Every Day)
2. Get Rows (n8n Table): Fetch all rows where updatedAt is recent (e.g., last 48 hours).
3. Loop over Items: Process one example at a time.
4. Delete Old Version (Supabase):
   Crucial Step Before adding the updated example, delete any existing row with the same ID to prevent duplicates.
5. Add Document (Supabase Vector Store):
   • Model: text-embedding-3-large
   • Dimensions: 1024
   • Map Data:
     • Page Content → requirements (This is what we embed).
     • Metadata → Map the full object (id, recommendation, explanation).

This is the live agent that responds to users.

1. Trigger: Chat or Webhook (Input: user_requirements)
2. Retrieve (Supabase Vector Store):
   • Operation: Get Many (Retrieve Ranked Documents).
   • Query: {{ $trigger.user_requirements }}
   • Limit: 4
   • Function Name: relevant_examples
3. Format Context (Code Node): Take the 4 JSON results and flatten them into a string.
4. LLM Chain (OpenAI GPT-4):

System Prompt:

You are a website recommendation expert.
To help you answer the user, here are examples of how we handled similar requests in the past:

{{ $json.formatted_examples }}

Analyze the user's request based on these precedents:
User Request: {{ $trigger.user_requirements }}