Self-Hosted Initialization

Configure TypeGraph in self-hosted mode with pgvector, separate ingest/search embedders, an LLM, and ontology config.

Install Packages

Self-hosted TypeGraph needs the SDK, the pgvector adapter, a database client, and whichever AI SDK provider you use for embeddings and extraction.

terminal

pnpm add @typegraph-ai/sdk @typegraph-ai/adapter-pgvector @ai-sdk/gateway @neondatabase/serverless

Build the Config

The end-state self-hosted config is tenantId + vectorStore + embedding + searchEmbedding + llm + ontology. embedding is used for ingest. searchEmbedding is optional but lets you use a separate search embedder in the same vector space.

typegraph.ts

import { gateway } from '@ai-sdk/gateway'
import { neon } from '@neondatabase/serverless'
import { PgVectorAdapter } from '@typegraph-ai/adapter-pgvector'
import { TenantId, typegraphDeploy, typegraphInit } from '@typegraph-ai/sdk'

const sql = neon(process.env.DATABASE_URL!)
const vectorStore = new PgVectorAdapter({ sql })

const config = {
  tenantId: TenantId('tenant_acme'),
  vectorStore,
  embedding: {
    model: gateway.embeddingModel('openai/text-embedding-3-small'),
    dimensions: 1536,
  },
  searchEmbedding: {
    model: gateway.embeddingModel('openai/text-embedding-3-small'),
    dimensions: 1536,
  },
  llm: {
    model: gateway.languageModel('openai/gpt-4.1-mini'),
  },
  ontology: {
    version: '2026-05-08',
    presets: ['b2b-saas'],
  },
}

Deploy Then Init

typegraphDeploy() provisions tables, extensions, and indexes. typegraphInit() creates the runtime client. Run deploy during setup or migrations; use init on normal application boot.

deploy-and-init.ts

await typegraphDeploy(config)

const typegraph = await typegraphInit(config)

Production Factory

Create one client per customer tenant. Tenant ID is the hard graph boundary; organization IDs from your app or Cloud auth should not be reused as the TypeGraph tenant unless that is truly your customer graph boundary.

factory.ts

export async function createTypeGraph(tenantId: string) {
  const sql = neon(process.env.DATABASE_URL!)
  const vectorStore = new PgVectorAdapter({ sql })

  const config = {
    tenantId: TenantId(tenantId),
    vectorStore,
    embedding: {
      model: gateway.embeddingModel('openai/text-embedding-3-small'),
      dimensions: 1536,
    },
    searchEmbedding: {
      model: gateway.embeddingModel('openai/text-embedding-3-small'),
      dimensions: 1536,
    },
    llm: {
      model: gateway.languageModel('openai/gpt-4.1-mini'),
    },
    ontology: {
      version: process.env.TYPEGRAPH_ONTOLOGY_VERSION ?? '2026-05-08',
      presets: ['b2b-saas'],
    },
  }

  return typegraphInit(config)
}

Optional Custom Extractor

If extractor is omitted, TypeGraph builds its default extractor from llm and ontology. Advanced self-hosted users can provide a custom extractor as long as it accepts the expected input and returns the expected extraction result.

custom-extractor.ts

const typegraph = await typegraphInit({
  ...config,
  extractor: {
    name: 'crm-extractor',
    capabilities: {
      supportsBatch: true,
      supportsStreaming: false,
      requiresStructuredOutput: false,
      multimodal: ['text'],
      preservesOntology: true,
    },
    async extract(input, ctx) {
      return {
        entities: [],
        facts: [],
        warnings: [],
        rawModelOutput: null,
      }
    },
  },
})

Back to All Guides

import { gateway } from '@ai-sdk/gateway' import { neon } from '@neondatabase/serverless' import { PgVectorAdapter } from '@typegraph-ai/adapter-pgvector' import { TenantId, typegraphDeploy, typegraphInit } from '@typegraph-ai/sdk' const sql = neon(process.env.DATABASE_URL!) const vectorStore = new PgVectorAdapter({ sql }) const config = { tenantId: TenantId('tenant_acme'), vectorStore, embedding: { model: gateway.embeddingModel('openai/text-embedding-3-small'), dimensions: 1536, }, searchEmbedding: { model: gateway.embeddingModel('openai/text-embedding-3-small'), dimensions: 1536, }, llm: { model: gateway.languageModel('openai/gpt-4.1-mini'), }, ontology: { version: '2026-05-08', presets: ['b2b-saas'], }, }

export async function createTypeGraph(tenantId: string) { const sql = neon(process.env.DATABASE_URL!) const vectorStore = new PgVectorAdapter({ sql }) const config = { tenantId: TenantId(tenantId), vectorStore, embedding: { model: gateway.embeddingModel('openai/text-embedding-3-small'), dimensions: 1536, }, searchEmbedding: { model: gateway.embeddingModel('openai/text-embedding-3-small'), dimensions: 1536, }, llm: { model: gateway.languageModel('openai/gpt-4.1-mini'), }, ontology: { version: process.env.TYPEGRAPH_ONTOLOGY_VERSION ?? '2026-05-08', presets: ['b2b-saas'], }, } return typegraphInit(config) }

const typegraph = await typegraphInit({ ...config, extractor: { name: 'crm-extractor', capabilities: { supportsBatch: true, supportsStreaming: false, requiresStructuredOutput: false, multimodal: ['text'], preservesOntology: true, }, async extract(input, ctx) { return { entities: [], facts: [], warnings: [], rawModelOutput: null, } }, }, })