DuckDB vector store

The DuckDB storage implementation provides an embedded high-performance vector search solution using DuckDB, an in-process analytical database. It uses the VSS extension for vector similarity search with HNSW indexing, offering a lightweight and efficient vector database that requires no external server.

It's part of the @mastra/duckdb package and offers efficient vector similarity search with metadata filtering.

Installation

npm

npm install @mastra/duckdb@latest

pnpm

pnpm add @mastra/duckdb@latest

Yarn

yarn add @mastra/duckdb@latest

Bun

bun add @mastra/duckdb@latest

Usage

import { DuckDBVector } from "@mastra/duckdb";

// Create a new vector store instance
const store = new DuckDBVector({
  id: "duckdb-vector",
  path: ":memory:", // or './vectors.duckdb' for file persistence
});

// Create an index
await store.createIndex({
  indexName: "myCollection",
  dimension: 1536,
  metric: "cosine",
});

// Add vectors with metadata
const vectors = [[0.1, 0.2, ...], [0.3, 0.4, ...]];
const metadata = [
  { text: "first document", category: "A" },
  { text: "second document", category: "B" },
];
await store.upsert({
  indexName: "myCollection",
  vectors,
  metadata,
});

// Query similar vectors
const queryVector = [0.1, 0.2, ...];
const results = await store.query({
  indexName: "myCollection",
  queryVector,
  topK: 10,
  filter: { category: "A" },
});

// Clean up
await store.close();

Constructor options

id:

string

Unique identifier for the vector store instance

path?:

string

= ':memory:'

Database file path. Use ':memory:' for in-memory database, or a file path like './vectors.duckdb' for persistence.

dimensions?:

number

= 1536

Default dimension for vector embeddings

metric?:

'cosine' | 'euclidean' | 'dotproduct'

= cosine

Default distance metric for similarity search

Methods

createIndex()

Creates a new vector collection with optional HNSW index for fast approximate nearest neighbor search.

indexName:

string

Name of the index to create

dimension:

number

Vector dimension size (must match your embedding model)

metric?:

'cosine' | 'euclidean' | 'dotproduct'

= cosine

Distance metric for similarity search

upsert()

Adds or updates vectors and their metadata in the index.

indexName:

string

Name of the index to insert into

vectors:

number[][]

Array of embedding vectors

metadata?:

Record<string, any>[]

Metadata for each vector

ids?:

string[]

Optional vector IDs (auto-generated UUIDs if not provided)

query()

Searches for similar vectors with optional metadata filtering.

indexName:

string

Name of the index to search in

queryVector:

number[]

Query vector to find similar vectors for

topK?:

number

= 10

Number of results to return

filter?:

Filter

Metadata filters using MongoDB-like query syntax

includeVector?:

boolean

= false

Whether to include vector data in results

describeIndex()

Gets information about an index.

indexName:

string

Name of the index to describe

Returns:

interface IndexStats {
  dimension: number
  count: number
  metric: 'cosine' | 'euclidean' | 'dotproduct'
}

deleteIndex()

Deletes an index and all its data.

indexName:

string

Name of the index to delete

listIndexes()

Lists all vector indexes in the database.

Returns: Promise<string[]>

updateVector()

Update a single vector by ID or by metadata filter. Either id or filter must be provided, but not both.

indexName:

string

Name of the index containing the vector

id?:

string

ID of the vector entry to update (mutually exclusive with filter)

filter?:

Record<string, any>

Metadata filter to identify vector(s) to update (mutually exclusive with id)

update:

object

Update data containing vector and/or metadata

update.vector?:

number[]

New vector data to update

update.metadata?:

Record<string, any>

New metadata to update

deleteVector()

Deletes a specific vector entry from an index by its ID.

indexName:

string

Name of the index containing the vector

id:

string

ID of the vector entry to delete

deleteVectors()

Delete multiple vectors by IDs or by metadata filter. Either ids or filter must be provided, but not both.

indexName:

string

Name of the index containing the vectors to delete

ids?:

string[]

Array of vector IDs to delete (mutually exclusive with filter)

filter?:

Record<string, any>

Metadata filter to identify vectors to delete (mutually exclusive with ids)

close()

Closes the database connection and releases resources.

await store.close()

Response types

Query results are returned in this format:

interface QueryResult {
  id: string
  score: number
  metadata: Record<string, any>
  vector?: number[] // Only included if includeVector is true
}

Filter operators

DuckDB vector store supports MongoDB-like filter operators:

Category	Operators
Comparison	$eq, $ne, $gt, $gte, $lt, $lte
Logical	$and, $or, $not, $nor
Array	$in, $nin, $all
Element	$exists
Text	$contains

Filter Examples

// Allegato operators
const results = await store.query({
  indexName: "docs",
  queryVector: [...],
  filter: {
    $and: [
      { category: "electronics" },
      { price: { $gte: 100, $lte: 500 } },
    ],
  },
});

// Nested field access
const results = await store.query({
  indexName: "docs",
  queryVector: [...],
  filter: { "user.profile.tier": "premium" },
});

Distance metrics

Metric	Description	Score Interpretation	Best For
cosine	Cosine similarity	0-1 (1 = most similar)	Text embeddings, normalized vectors
euclidean	L2 distance	0-∞ (0 = most similar)	Image embeddings, spatial data
dotproduct	Inner product	Higher = more similar	When vector magnitude matters

Error handling

The store throws specific errors for different failure cases:

try {
  await store.query({
    indexName: 'my-collection',
    queryVector: queryVector,
  })
} catch (error) {
  if (error.message.includes('not found')) {
    console.error('The specified index does not exist')
  } else if (error.message.includes('Invalid identifier')) {
    console.error('Index name contains invalid characters')
  } else {
    console.error('Vector store error:', error.message)
  }
}

Common error cases include:

• Invalid index name format
• Index/table not found
• Dimension mismatch between query vector and index
• Empty filter or ids array in delete/update operations
• Mutual exclusivity violations (providing both id and filter)

Use cases

Embedded Semantic Search

Build offline-capable AI applications with semantic search that runs entirely in-process:

const store = new DuckDBVector({
  id: 'offline-search',
  path: './search.duckdb',
})

Local RAG Pipelines

Process sensitive documents locally without sending data to cloud vector databases:

const store = new DuckDBVector({
  id: 'private-rag',
  path: './confidential.duckdb',
  dimensions: 1536,
})

Development and Testing

Rapidly prototype vector search features with zero infrastructure:

const store = new DuckDBVector({
  id: 'dev-store',
  path: ':memory:', // Fast in-memory for tests
})

Related

• Metadata Filters
• DuckDB Documentation