Graphile + pgvector Integration

Integrate pgvector with PostGraphile v5 (Grafast/graphile-build-pg) so that vector(n) columns are surfaced in GraphQL as a Vector scalar, and vector similarity search is available as first-class GraphQL query fields.

When to Apply

Use this skill when:

• Adding vector(n) columns to tables and need them visible in the PostGraphile schema
• Exposing vector similarity search (cosine, L2, inner product) as GraphQL queries
• Building AI/RAG features on top of a Constructive or PostGraphile v5 stack
• Writing integration tests for pgvector-backed GraphQL APIs

Two Moving Parts

Both must be in the preset for end-to-end vector support.

VectorCodecPlugin — Expose vector Columns as GraphQL Scalar

Without this plugin, PostGraphile silently ignores vector(n) columns. The codec teaches graphile-build-pg how to parse and serialize the type.

Wire Format

PostgreSQL sends vectors as text: [0.1,0.2,...,0.768] JavaScript representation: number[] GraphQL scalar: Vector (serialized as [Float])

Source Location

graphile-settings/src/plugins/vector-codec.ts — included automatically in ConstructivePreset.

Pure Helper Exports (Testable Standalone)

import { fromPgVector, toPgVector, vectorScalarConfig } from 'graphile-settings/plugins/vector-codec';

// Parse PG text → JS
fromPgVector('[0.1,0.2,0.3]'); // => [0.1, 0.2, 0.3]

// Serialize JS → PG text
toPgVector([0.1, 0.2, 0.3]); // => '[0.1,0.2,0.3]'

Plugin Structure

export const VectorCodecPlugin: GraphileConfig.Plugin = {
  name: 'VectorCodecPlugin',
  version: '1.0.0',

  // Gather phase: hook into codec discovery; register codec when type.typname === 'vector'
  gather: {
    hooks: {
      async pgCodecs_findPgCodec(info, event) {
        if (event.pgCodec) return;
        const { pgType: type, serviceName } = event;
        if (type.typname !== 'vector') return;
        // ... resolve namespace, build codec with fromPg/toPg, set event.pgCodec
      },
    },
  },

  // Schema phase: register Vector scalar and map the codec to it
  schema: {
    hooks: {
      init: {
        before: ['PgCodecs'],
        callback(_, build) {
          build.registerScalarType('Vector', {}, () => vectorScalarConfig, '...');
          for (const codec of Object.values(build.input.pgRegistry.pgCodecs)) {
            if ((codec as any).name === 'vector') {
              setGraphQLTypeForPgCodec(codec as any, 'input', 'Vector');
              setGraphQLTypeForPgCodec(codec as any, 'output', 'Vector');
            }
          }
          return _;
        },
      },
    },
  },
};

Key Hook: pgCodecs_findPgCodec

This gather hook fires for every unknown PG type. Return early if event.pgCodec is already set (already handled). Check type.typname === 'vector' and assign a codec with fromPg/toPg functions.

Key Hook: init (before PgCodecs)

Must run before PgCodecs processes the registry. Registers the Vector scalar and maps it to the codec for both input and output GraphQL roles.

PgVectorPlugin — Vector Search GraphQL Fields

Adds root query fields for similarity search across configured tables.

Installation

pnpm add postgraphile-plugin-pgvector

Usage — Preset (Recommended)

import { PgVectorPreset } from 'postgraphile-plugin-pgvector';

const preset: GraphileConfig.Preset = {
  extends: [
    ConstructivePreset,        // includes VectorCodecPlugin
    PgVectorPreset({
      collections: [
        {
          schema: 'public',
          table: 'documents',
          embeddingColumn: 'embedding',
          graphqlFieldName: 'vectorSearchDocument',  // optional; auto-generated if omitted
          maxQueryDim: 768,                          // optional defensive validation
        },
        {
          schema: 'public',
          table: 'contacts',
          embeddingColumn: 'embedding',
        },
      ],
      defaultMetric: 'COSINE',  // COSINE | L2 | IP
      maxLimit: 100,
    }),
  ],
};

Usage — Plugin Directly

import { PgVectorPlugin } from 'postgraphile-plugin-pgvector';

const plugin = PgVectorPlugin({
  collections: [{ schema: 'public', table: 'documents', embeddingColumn: 'embedding' }],
});

Generated GraphQL API

type Query {
  vectorSearchDocument(
    query: [Float!]!       # embedding vector (must match column dimensions)
    limit: Int             # default 10, capped at maxLimit
    offset: Int            # pagination
    metric: VectorMetric   # COSINE (default), L2, IP
  ): [DocumentVectorSearchResult!]!
}

enum VectorMetric { COSINE L2 IP }

type DocumentVectorSearchResult {
  distance: Float!   # distance score; lower = more similar for COSINE/L2
  id: Int!
  title: String!
  # ... all other table columns
}

GraphQL Query Example

query SearchDocuments($vector: [Float!]!) {
  vectorSearchDocument(query: $vector, limit: 5, metric: COSINE) {
    distance
    id
    title
    content
  }
}

Distance Metrics

Codegen — Vector Type → number[]

When using @constructive-io/graphql-codegen, add the Vector scalar mapping so generated types use number[] instead of unknown:

// graphql-codegen.config.ts
export default createConfig({
  targets: [{
    // ...
    scalars: {
      Vector: 'number[]',
    },
  }],
});

With @constructive-io/graphql-codegen >= 4.6.0, Vector: 'number[]' is built-in to SCALAR_TS_MAP. For 4.2.0, apply a pnpm patch adding scalars?: Record<string, string> to the config type defs.

SQL Search Functions (Alternative: Direct SQL)

For route-level RAG (bypassing GraphQL), define SQL functions instead:

CREATE OR REPLACE FUNCTION your_schema.search_contacts(
  query_embedding vector(768),
  match_count int DEFAULT 10
)
RETURNS TABLE (
  id uuid,
  entity_id text,
  first_name text,
  last_name text,
  email text,
  headline text,
  similarity float
)
LANGUAGE sql STABLE
AS $$
  SELECT
    id, entity_id, first_name, last_name, email, headline,
    1 - (embedding <=> query_embedding) AS similarity
  FROM your_schema.contacts
  WHERE embedding IS NOT NULL
  ORDER BY embedding <=> query_embedding
  LIMIT match_count;
$$;

These are surfaced automatically in the GraphQL schema (PostGraphile computed columns / functions) if the naming and return type conventions match. The VectorCodecPlugin is still required so that the vector argument type resolves correctly.

Integration Tests

Use graphile-test's getConnections with real DB. No mocks.

// __tests__/pgvector.test.ts
import { join } from 'path';
import { getConnections, seed } from 'graphile-test';
import { PgVectorPreset } from 'postgraphile-plugin-pgvector';
import { ConstructivePreset } from 'graphile-settings';

describe('PgVectorPlugin', () => {
  let db: PgTestClient;
  let teardown: () => Promise<void>;
  let query: QueryFn;

  beforeAll(async () => {
    const testPreset = {
      extends: [
        ConstructivePreset,
        PgVectorPreset({
          collections: [{
            schema: 'pgvector_test',
            table: 'documents',
            embeddingColumn: 'embedding',
            graphqlFieldName: 'vectorSearchDocument',
          }],
        }),
      ],
    };

    const connections = await getConnections({
      schemas: ['pgvector_test'],
      preset: testPreset,
      useRoot: true,
    }, [
      seed.sqlfile([join(__dirname, './setup.sql')])
    ]);

    ({ db, teardown, query } = connections);
    await db.client.query('BEGIN');
  });

  afterAll(async () => {
    try { await db.client.query('ROLLBACK'); } catch {}
    await teardown();
  });

  beforeEach(() => db.beforeEach());
  afterEach(() => db.afterEach());

  it('returns vector search results ordered by distance', async () => {
    // Insert test doc with pre-computed embedding
    const embedding = Array(768).fill(0.1);
    await db.client.query(
      `INSERT INTO pgvector_test.documents (title, content, embedding)
       VALUES ($1, $2, $3::vector)`,
      ['Test Doc', 'Hello world', `[${embedding.join(',')}]`]
    );

    const result = await query<{ vectorSearchDocument: Array<{ title: string; distance: number }> }>(`
      query {
        vectorSearchDocument(query: [${embedding.join(',')}], limit: 5, metric: COSINE) {
          title
          distance
        }
      }
    `);

    expect(result.data?.vectorSearchDocument).toHaveLength(1);
    expect(result.data?.vectorSearchDocument[0].title).toBe('Test Doc');
    expect(result.data?.vectorSearchDocument[0].distance).toBeCloseTo(0, 5);
  });
});

Setup SQL

-- __tests__/setup.sql
CREATE SCHEMA IF NOT EXISTS pgvector_test;
CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE pgvector_test.documents (
  id SERIAL PRIMARY KEY,
  title TEXT NOT NULL,
  content TEXT,
  embedding vector(768),
  created_at TIMESTAMPTZ DEFAULT now()
);

CREATE INDEX ON pgvector_test.documents USING hnsw (embedding vector_cosine_ops);

Key graphile-test Pattern Notes

• getConnections({ schemas, preset, useRoot: true }, [seed.sqlfile([...])]) — no mocks
• graphile-test has MinimalPreset baked in; pass your plugin preset as extends
• SQL fixture must CREATE EXTENSION IF NOT EXISTS vector; before creating vector columns
• Use BEGIN / ROLLBACK for transaction-scoped test isolation (not beforeEach schema drops)

Codec-Only Usage (No Search Fields)

If you only need vector columns visible in the schema (not search fields), just include VectorCodecPlugin in your preset:

import { VectorCodecPlugin } from 'graphile-settings/plugins/vector-codec';
// or use ConstructivePreset which includes it automatically

const preset = {
  plugins: [VectorCodecPlugin],
};

After adding: embedding: number[] | null appears in all generated types, mutations accept Vector scalar, and SQL functions with vector args are surfaced in GraphQL.

Common Pitfalls

References

• Related skill: pgvector-setup — DB schema with HNSW indexes
• Related skill: pgvector-embeddings — generating and storing embeddings
• Related skill: rag-pipeline — full RAG pipeline from embed to LLM
• Related skill: agentic-kit-rag — wiring RAG into agentic-kit chat routes
• pgvector docs
• PostGraphile v5 plugin guide
• Grafast step plan API