Give your chatbot memory of your docs with Supabase vector search

What you get

Your own documents stored as embeddings in Supabase Postgres with pgvector, plus a search script that finds the passages closest in meaning to a question, ready to feed into any chatbot prompt.

Prerequisites

A Supabase project (the Free plan works)
Node.js 18 or later
Your project URL and a secret key (sb_secret_...) from the project's API keys settings. The scripts run server-side only; never ship a secret key to a browser.

Steps

In the Supabase dashboard, open the SQL Editor and enable pgvector:

SQL

create extension vector
with
  schema extensions;

In the SQL Editor, create a table for your documents. The gte-small model outputs 384 dimensions, so the vector column is vector(384):

SQL

create table documents (
  id serial primary key,
  title text not null,
  body text not null,
  embedding extensions.vector(384)
);

In the SQL Editor, create the match_documents function that ranks rows by cosine similarity:

SQL

create or replace function match_documents (
  query_embedding extensions.vector(384),
  match_threshold float,
  match_count int
)
returns table (
  id bigint,
  title text,
  body text,
  similarity float
)
language sql stable
as $$
  select
    documents.id,
    documents.title,
    documents.body,
    1 - (documents.embedding <=> query_embedding) as similarity
  from documents
  where 1 - (documents.embedding <=> query_embedding) > match_threshold
  order by (documents.embedding <=> query_embedding) asc
  limit match_count;
$$;

In a new folder, install the Supabase client and Transformers.js (which runs the open gte-small embedding model locally, no API key needed):

Shell

npm install @supabase/supabase-js @huggingface/transformers

Create embed.mjs. It embeds each document and stores it in the documents table (replace the sample docs with your own content):

JavaScript

import { createClient } from '@supabase/supabase-js'
import { pipeline } from '@huggingface/transformers'

const supabase = createClient(process.env.SUPABASE_URL, process.env.SUPABASE_SECRET_KEY)

const generateEmbedding = await pipeline('feature-extraction', 'Supabase/gte-small')

const docs = [
  { title: 'Refunds', body: 'You can request a full refund within 30 days of purchase.' },
  { title: 'Shipping', body: 'Orders ship within 2 business days from our warehouse.' },
  { title: 'Support hours', body: 'Our support team replies Monday to Friday, 9am to 5pm.' },
]

for (const { title, body } of docs) {
  const output = await generateEmbedding(body, {
    pooling: 'mean',
    normalize: true,
  })

  const embedding = Array.from(output.data)

  const { error } = await supabase.from('documents').insert({
    title,
    body,
    embedding,
  })

  if (error) throw error
  console.log('Stored:', title)
}

Create search.mjs. It embeds a question with the same model and calls match_documents:

JavaScript

import { createClient } from '@supabase/supabase-js'
import { pipeline } from '@huggingface/transformers'

const supabase = createClient(process.env.SUPABASE_URL, process.env.SUPABASE_SECRET_KEY)

const generateEmbedding = await pipeline('feature-extraction', 'Supabase/gte-small')

const query = process.argv[2] ?? 'How long do I have to get my money back?'

const output = await generateEmbedding(query, {
  pooling: 'mean',
  normalize: true,
})

const embedding = Array.from(output.data)

const { data: documents, error } = await supabase.rpc('match_documents', {
  query_embedding: embedding,
  match_threshold: 0.78,
  match_count: 10,
})

if (error) throw error
console.log(documents)

Set SUPABASE_URL and SUPABASE_SECRET_KEY in your environment, then store your documents:

Shell

node embed.mjs

Test it

Run a search with a question that shares no keywords with the stored text:

Shell

node search.mjs "How long do I have to get my money back?"

You should get the Refunds row back with a similarity score above 0.78. If nothing is returned, lower match_threshold (for example to 0.5). To give a chatbot memory of your docs, pass the returned body values into your model prompt as context.

What you'll need

What you get

Prerequisites

Steps

Test it