AI & Vectors

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:

  1. Go to the SQL Editor page in the Dashboard.
  2. Click LangChain in the Quick start section.
  3. Click Run.

Usage

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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
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_KEYif (!supabaseKey) throw new Error(`Expected SUPABASE_SERVICE_ROLE_KEY`)const url = process.env.SUPABASE_URLif (!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.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
import { SupabaseVectorStore } from 'langchain/vectorstores/supabase'import { OpenAIEmbeddings } from 'langchain/embeddings/openai'import { createClient } from '@supabase/supabase-js'// First, follow set-up instructions aboveconst privateKey = process.env.SUPABASE_SERVICE_ROLE_KEYif (!privateKey) throw new Error(`Expected env var SUPABASE_SERVICE_ROLE_KEY`)const url = process.env.SUPABASE_URLif (!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).

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
import { SupabaseFilterRPCCall, SupabaseVectorStore } from 'langchain/vectorstores/supabase'import { OpenAIEmbeddings } from 'langchain/embeddings/openai'import { createClient } from '@supabase/supabase-js'// First, follow set-up instructions aboveconst privateKey = process.env.SUPABASE_SERVICE_ROLE_KEYif (!privateKey) throw new Error(`Expected env var SUPABASE_SERVICE_ROLE_KEY`)const url = process.env.SUPABASE_URLif (!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 theoretical 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)}

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