@typegraph-ai/adapter-pgvector

PostgreSQL + pgvector adapter for TypeGraph. Hybrid search combining vector similarity with keyword matching via tsvector and Reciprocal Rank Fusion.

Bring your own Postgres driver -- works with Neon serverless, node-postgres, Drizzle, or anything that can run a parameterized query. Per-model table isolation keeps embedding dimensions and indexes clean.

Install

npm install @typegraph-ai/adapter-pgvector @typegraph-ai/core

Usage

import { neon } from '@neondatabase/serverless'
import { PgVectorAdapter } from '@typegraph-ai/adapter-pgvector'
import { typegraph } from '@typegraph-ai/core'

const sql = neon(process.env.DATABASE_URL!)

const adapter = new PgVectorAdapter({ sql })

const agent = await typegraph.initialize({
  adapter,
  // ... embedding provider, etc.
})

The SqlExecutor pattern means you wire up the driver yourself:

// node-postgres
import { Pool } from 'pg'
const pool = new Pool({ connectionString: '...' })
const sql: SqlExecutor = (q, p) => pool.query(q, p).then(r => r.rows)

Exports

| Export | Description | |--------|-------------| | PgVectorAdapter | Main adapter class, implements VectorStoreAdapter | | PgHashStore | Content-hash deduplication store | | PgDocumentStore | Document record CRUD | | REGISTRY_SQL | DDL for model registry table | | MODEL_TABLE_SQL | DDL for per-model chunk tables | | HASH_TABLE_SQL | DDL for hash deduplication table | | DOCUMENTS_TABLE_SQL | DDL for documents table | | BUCKETS_TABLE_SQL | DDL for buckets table | | JOBS_TABLE_SQL | DDL for jobs table | | sanitizeModelKey | Normalizes model names into safe table suffixes |

Types

| Type | Description | |------|-------------| | PgVectorAdapterConfig | Constructor options (sql, transaction, schema, tablePrefix, hashesTable, documentsTable) | | SqlExecutor | (query: string, params?: unknown[]) => Promise<Record<string, unknown>[]> |

Related

• TypeGraph main repo
• Self-Hosted Setup Guide