How to Configure Prisma ORM with pgvector for Vector Search

1. Enable the pgvector extension

First, ensure your PostgreSQL instance has the pgvector extension available, then enable it in a migration.

-- migration.sql
CREATE EXTENSION IF NOT EXISTS vector;

-- table for embedded documents
CREATE TABLE "DocumentChunk" (
  "id" SERIAL PRIMARY KEY,
  "content" TEXT NOT NULL,
  "embedding" vector(1536) NOT NULL,
  "documentId" INTEGER NOT NULL
);

2. Model it in the Prisma schema

Prisma does not have a native vector type, so store embeddings as a raw type and manage them via raw SQL for inserts/queries.

model DocumentChunk {
  id        Int    @id @default(autoincrement())
  content   String
  embedding Unsupported("vector(1536)")?
  documentId Int
}

3. Insert and query with cosine similarity

Insert embeddings as a parameterized vector literal, then order by the `<=>` cosine-distance operator (smaller = closer).

import { PrismaClient } from "@prisma/client";
const prisma = new PrismaClient();

// insert
await prisma.$executeRaw
  `INSERT INTO "DocumentChunk" ("content","embedding","documentId")
    VALUES (${content}, ${emb}::vector, ${docId})`;

// nearest neighbors
const hits = await prisma.$queryRaw
  `SELECT "content", "embedding" <=> ${emb}::vector AS distance
     FROM "DocumentChunk" ORDER BY distance LIMIT 5`;

4. Performance notes

Add an IVFFlat or HNSW index on the embedding column for large datasets.

Keep embedding dimensions consistent with your model (e.g., 1536 for OpenAI text-embedding-3-small).

FAQ