LangChain

LangChain is a popular framework for working with AI, Vectors, and embeddings. LangChain supports using Supabase as a vector store, using the pgvector extension.

Initializing your database#

Prepare you database with the relevant tables:

-- Enable the pgvector extension to work with embedding vectors
create extension vector;
-- Create a table to store your documents
create table documents (
  id bigserial primary key,
  content text, -- corresponds to Document.pageContent
  metadata jsonb, -- corresponds to Document.metadata
  embedding vector(1536) -- 1536 works for OpenAI embeddings, change if needed
);
-- Create a function to search for documents
create function match_documents (
  query_embedding vector(1536),
  match_count int default null,
  filter jsonb DEFAULT '{}'
) returns table (
  id bigint,
  content text,
  metadata jsonb,
  similarity float
)
language plpgsql
as $$
#variable_conflict use_column
begin
  return query
  select
    id,
    content,
    metadata,
    1 - (documents.embedding <=> query_embedding) as similarity
  from documents
  where metadata @> filter
  order by documents.embedding <=> query_embedding
  limit match_count;
end;
$$;

Usage#

You can now search your documents using any Node.js application. This is intended to be run on a secure server route.

import { SupabaseVectorStore } from 'langchain/vectorstores/supabase'
import { OpenAIEmbeddings } from 'langchain/embeddings/openai'
import { createClient } from '@supabase/supabase-js'
const supabaseKey = process.env.SUPABASE_SERVICE_ROLE_KEY
if (!supabaseKey) throw new Error(`Expected SUPABASE_SERVICE_ROLE_KEY`)
const url = process.env.SUPABASE_URL
if (!url) throw new Error(`Expected env var SUPABASE_URL`)
export const run = async () => {
  const client = createClient(url, supabaseKey)
  const vectorStore = await SupabaseVectorStore.fromTexts(
    ['Hello world', 'Bye bye', "What's this?"],
    [{ id: 2 }, { id: 1 }, { id: 3 }],
    new OpenAIEmbeddings(),
    {
      client,
      tableName: 'documents',
      queryName: 'match_documents',
    }
  )
  const resultOne = await vectorStore.similaritySearch('Hello world', 1)
  console.log(resultOne)
}

Simple Metadata Filtering#

Given the above match_documents Postgres function, you can also pass a filter parameter to only return documents with a specific metadata field value. This filter parameter is a JSON object, and the match_documents function will use the Postgres JSONB Containment operator @> to filter documents by the metadata field values you specify. See details on the Postgres JSONB Containment operator for more information.

import { SupabaseVectorStore } from 'langchain/vectorstores/supabase'
import { OpenAIEmbeddings } from 'langchain/embeddings/openai'
import { createClient } from '@supabase/supabase-js'
// First, follow set-up instructions above
const privateKey = process.env.SUPABASE_SERVICE_ROLE_KEY
if (!privateKey) throw new Error(`Expected env var SUPABASE_SERVICE_ROLE_KEY`)
const url = process.env.SUPABASE_URL
if (!url) throw new Error(`Expected env var SUPABASE_URL`)
export const run = async () => {
  const client = createClient(url, privateKey)
  const vectorStore = await SupabaseVectorStore.fromTexts(
    ['Hello world', 'Hello world', 'Hello world'],
    [{ user_id: 2 }, { user_id: 1 }, { user_id: 3 }],
    new OpenAIEmbeddings(),
    {
      client,
      tableName: 'documents',
      queryName: 'match_documents',
    }
  )
  const result = await vectorStore.similaritySearch('Hello world', 1, {
    user_id: 3,
  })
  console.log(result)
}

Advanced Metadata Filtering#

You can also use query builder-style filtering (similar to how the Supabase JavaScript library works) instead of passing an object. Note that since the filter properties will be in the metadata column, you need to use arrow operators (-> for integer or ->> for text) as defined in Postgrest API documentation and specify the data type of the property (e.g. the column should look something like metadata->some_int_value::int).

import { SupabaseFilterRPCCall, SupabaseVectorStore } from 'langchain/vectorstores/supabase'
import { OpenAIEmbeddings } from 'langchain/embeddings/openai'
import { createClient } from '@supabase/supabase-js'
// First, follow set-up instructions above
const privateKey = process.env.SUPABASE_SERVICE_ROLE_KEY
if (!privateKey) throw new Error(`Expected env var SUPABASE_SERVICE_ROLE_KEY`)
const url = process.env.SUPABASE_URL
if (!url) throw new Error(`Expected env var SUPABASE_URL`)
export const run = async () => {
  const client = createClient(url, privateKey)
  const embeddings = new OpenAIEmbeddings()
  const store = new SupabaseVectorStore(embeddings, {
    client,
    tableName: 'documents',
  })
  const docs = [
    {
      pageContent:
        'This is a long text, but it actually means something because vector database does not understand Lorem Ipsum. So I would need to expand upon the notion of quantum fluff, a theorectical concept where subatomic particles coalesce to form transient multidimensional spaces. Yet, this abstraction holds no real-world application or comprehensible meaning, reflecting a cosmic puzzle.',
      metadata: { b: 1, c: 10, stuff: 'right' },
    },
    {
      pageContent:
        'This is a long text, but it actually means something because vector database does not understand Lorem Ipsum. So I would need to proceed by discussing the echo of virtual tweets in the binary corridors of the digital universe. Each tweet, like a pixelated canary, hums in an unseen frequency, a fascinatingly perplexing phenomenon that, while conjuring vivid imagery, lacks any concrete implication or real-world relevance, portraying a paradox of multidimensional spaces in the age of cyber folklore.',
      metadata: { b: 2, c: 9, stuff: 'right' },
    },
    { pageContent: 'hello', metadata: { b: 1, c: 9, stuff: 'right' } },
    { pageContent: 'hello', metadata: { b: 1, c: 9, stuff: 'wrong' } },
    { pageContent: 'hi', metadata: { b: 2, c: 8, stuff: 'right' } },
    { pageContent: 'bye', metadata: { b: 3, c: 7, stuff: 'right' } },
    { pageContent: "what's this", metadata: { b: 4, c: 6, stuff: 'right' } },
  ]
  await store.addDocuments(docs)
  const funcFilterA: SupabaseFilterRPCCall = (rpc) =>
    rpc
      .filter('metadata->b::int', 'lt', 3)
      .filter('metadata->c::int', 'gt', 7)
      .textSearch('content', `'multidimensional' & 'spaces'`, {
        config: 'english',
      })
  const resultA = await store.similaritySearch('quantum', 4, funcFilterA)
  const funcFilterB: SupabaseFilterRPCCall = (rpc) =>
    rpc
      .filter('metadata->b::int', 'lt', 3)
      .filter('metadata->c::int', 'gt', 7)
      .filter('metadata->>stuff', 'eq', 'right')
  const resultB = await store.similaritySearch('hello', 2, funcFilterB)
  console.log(resultA, resultB)
}

Hybrid search#

LangChain supports the concept of a hybrid search, which combines Similarity Search with Full Text Search. Read the official docs to get started: Supabase Hybrid Search.

You can install the LangChain Hybrid Search function though our database.dev package manager.

Resources#

Official LangChain site.
Official LangChain docs.
Supabase Hybrid Search.

Edit this page on GitHub