timescale/pgvectorscale

[!NOTE] A new approach to vector search: We're building pg_vectorsearch, a next-generation PostgreSQL vector index that uses partitioning and compression instead of graph-based indexing to deliver faster searches with less memory. It's available in early alpha and isn't yet production-ready. Learn more.

pgvectorscale complements pgvector, the open-source vector data extension for PostgreSQL, and introduces the following key innovations for pgvector data:

• A new index type called StreamingDiskANN, inspired by the DiskANN algorithm, based on research from Microsoft.
• Statistical Binary Quantization: developed by Timescale researchers, This compression method improves on standard Binary Quantization.
• Label-based filtered vector search: based on Microsoft's Filtered DiskANN research, this allows you to combine vector similarity search with label filtering for more precise and efficient results.

On a benchmark dataset of 50 million Cohere embeddings with 768 dimensions each, PostgreSQL with pgvector and pgvectorscale achieves 28x lower p95 latency and 16x higher query throughput compared to Pinecone's storage optimized (s1) index for approximate nearest neighbor queries at 99% recall, all at 75% less cost when self-hosted on AWS EC2.

To learn more about the performance impact of pgvectorscale, and details about benchmark methodology and results, see the pgvector vs Pinecone comparison blog post.

In contrast to pgvector, which is written in C, pgvectorscale is developed in Rust using the PGRX framework, offering the PostgreSQL community a new avenue for contributing to vector support.

Application developers or DBAs can use pgvectorscale with their PostgreSQL databases.

• Install pgvectorscale
• Get started using pgvectorscale

If you want to contribute to this extension, see how to build pgvectorscale from source in a developer environment and our testing guide.

For production vector workloads, get private beta access to vector-optimized databases with pgvector and pgvectorscale on Timescale. Sign up here for priority access.

Installation

The fastest ways to run PostgreSQL with pgvectorscale are:

• Using a pre-built Docker container
• Installing from source
• Enable pgvectorscale in a Timescale Cloud service

Using a pre-built Docker container

1. Run the TimescaleDB Docker image.

2. Connect to your database:

   psql -d "postgres://<username>:<password>@<host>:<port>/<database-name>"
   

3. Create the pgvectorscale extension:

   CREATE EXTENSION IF NOT EXISTS vectorscale CASCADE;
   

   The CASCADE automatically installs pgvector.

Installing from source

You can install pgvectorscale from source and install it in an existing PostgreSQL server

[!WARNING] Building pgvectorscale on macOS X86 (Intel) machines is currently not supported due to an open issue. As alternatives, you can:

• Use an ARM-based Mac.
• Build using Linux.
• Use our pre-built Docker containers.

We welcome community contributions to resolve this limitation. If you're interested in helping, please check the issue for details.

1. Compile and install the extension

   # install rust
   curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
   
   # download pgvectorscale
   cd /tmp
   git clone --branch <version> https://github.com/timescale/pgvectorscale
   cd pgvectorscale/pgvectorscale
   # install cargo-pgrx with the same version as pgrx
   cargo install --locked cargo-pgrx --version $(cargo metadata --format-version 1 | jq -r '.packages[] | select(.name == "pgrx") | .version')
   cargo pgrx init --pg18 pg_config
   # build and install pgvectorscale
   cargo pgrx install --release
   

   You can also take a look at our documentation for extension developers for more complete instructions.

2. Connect to your database:

   psql -d "postgres://<username>:<password>@<host>:<port>/<database-name>"
   

3. Ensure the pgvector extension is available:

   SELECT * FROM pg_available_extensions WHERE name = 'vector';
   

   If pgvector is not available, install it using the pgvector installation instructions.

4. Create the pgvectorscale extension:

   CREATE EXTENSION IF NOT EXISTS vectorscale CASCADE;
   

   The CASCADE automatically installs pgvector.

Enable pgvectorscale in a Timescale Cloud service

Note: the instructions below are for Timescale's standard compute instance. For production vector workloads, we're offering private beta access to vector-optimized databases with pgvector and pgvectorscale on Timescale. Sign up here for priority access.

To enable pgvectorscale:

1. Create a new Timescale Service.

   If you want to use an existing service, pgvectorscale is added as an available extension on the first maintenance window after the pgvectorscale release date.

2. Connect to your Timescale service:

   psql -d "postgres://<username>:<password>@<host>:<port>/<database-name>"
   

3. Create the pgvectorscale extension:

   CREATE EXTENSION IF NOT EXISTS vectorscale CASCADE;
   

   The CASCADE automatically installs pgvector.

Get started with pgvectorscale

1. Create a table with an embedding column. For example:

   CREATE TABLE IF NOT EXISTS document_embedding  (
       id BIGINT PRIMARY KEY GENERATED BY DEFAULT AS IDENTITY,
       metadata JSONB,
       contents TEXT,
       embedding VECTOR(1536)
   )
   

2. Populate the table.

   For more information, see the pgvector instructions and list of clients.

3. Create a StreamingDiskANN index on the embedding column:

   CREATE INDEX document_embedding_idx ON document_embedding
   USING diskann (embedding vector_cosine_ops);
   

4. Find the 10 closest embeddings using the index.

   SELECT *
   FROM document_embedding
   ORDER BY embedding <=> $1
   LIMIT 10;
   

   Note: pgvectorscale currently supports: cosine distance (<=>) queries, for indices created with vector_cosine_ops; L2 distance (<->) queries, for indices created with vector_l2_ops; and inner product (<#>) queries, for indices created with vector_ip_ops. This is the same syntax used by pgvector. If you would like additional distance types, create an issue. (Note: inner product indices are not compatible with plain storage.)

Filtered Vector Search

pgvectorscale supports combining vector similarity search with metadata filtering. There are two basic kinds of filtering, which can be combined in a single query:

1. Label-based filtering with the diskann index: This provides optimized performance for filtering by labels.
2. Arbitrary WHERE clause filtering: This uses post-filtering after the vector search.

The label-based filtering implementation is based on the Filtered DiskANN approach developed by Microsoft researchers, which enables efficient filtered vector search while maintaining high recall.

The post-filtering implementation, while slower, is streaming and correct, ensuring accurate results without requiring the entire result set to be loaded into memory.

Label-based Filtering with diskann

For optimal performance with label filtering, you must specify the label column directly in the index creation:

1. Create a table with an embedding column and a labels array:

   CREATE TABLE documents (
       id SERIAL PRIMARY KEY,
       embedding VECTOR(1536),
       labels SMALLINT[],  -- Array of category labels
       status TEXT,
       created_at TIMESTAMPTZ
   );
   

2. Create a StreamingDiskANN index on the embedding column, including the labels column:

   CREATE INDEX ON documents USING diskann (embedding vector_cosine_ops, labels);
   

Note: Label values must be within the PostgreSQL smallint range (-32768 to 32767). Using smallint[] for labels ensures that PostgreSQL's type system will automatically enforce these bounds.

pgvectorscale includes an implementation of the && overlap operator for smallint[] arrays, which is used for efficient label-based filtering.

3. Perform label-filtered vector searches using the && operator (array overlap):

   -- Find similar documents with specific labels
   SELECT * FROM documents
   WHERE labels && ARRAY[1, 3]  -- Documents with label 1 OR 3
   ORDER BY embedding <=> '[...]'
   LIMIT 10;
   

   The index directly supports this type of filtering, providing significantly lower latency results compared to post-filtering.

Giving Semantic Meaning to Labels

While the labels must be stored as integers in the array for the index to work efficiently, you can give them semantic meaning by relating them to a separate labels table:

1. Create a labels table with meaningful descriptions:

   CREATE TABLE label_definitions (
       id INTEGER PRIMARY KEY,
       name TEXT,
       description TEXT,
       attributes JSONB  -- Can store additional metadata about the label
   );
   
   -- Insert some label definitions
   INSERT INTO label_definitions (id, name, description, attributes) VALUES
   (1, 'science', 'Scientific content', '{"domain": "academic", "confidence": 0.95}'),
   (2, 'technology', 'Technology-related content', '{"domain": "technical", "confidence": 0.92}'),
   (3, 'business', 'Business and finance content', '{"domain": "commercial", "confidence": 0.88}');
   

2. When inserting documents, use the appropriate label IDs:

   -- Insert a document with science and technology labels
   INSERT INTO documents (embedding, labels)
   VALUES ('[...]', ARRAY[1, 2]);
   

3. When querying, you can join with the labels table to work with meaningful names:

   -- Find similar science documents and include label information
   SELECT d.*, array_agg(l.name) as label_names
   FROM documents d
   JOIN label_definitions l ON l.id = ANY(d.labels)
   WHERE d.labels && ARRAY[1]  -- Science label
   GROUP BY d.id, d.embedding, d.labels, d.status, d.created_at
   ORDER BY d.embedding <=> '[...]'
   LIMIT 10;
   

4. You can also convert between label names and IDs when filtering:

   -- Find documents with specific label names
   SELECT d.*
   FROM documents d
   WHERE d.labels && (
       SELECT array_agg(id)
       FROM label_definitions
       WHERE name IN ('science', 'business')
   )
   ORDER BY d.embedding <=> '[...]'
   LIMIT 10;
   

This approach gives you the performance benefits of integer-based label filtering while still allowing you to work with semantically meaningful labels in your application.

Arbitrary WHERE Clause Filtering

You can also use any PostgreSQL WHERE clause with vector search, but these conditions will be applied as post-filtering:

-- Find similar documents with specific status and date range
SELECT * FROM documents
WHERE status = 'active' AND created_at > '2024-01-01'
ORDER BY embedding <=> '[...]'
LIMIT 10;

For these arbitrary conditions, the vector search happens first, and then the WHERE conditions are applied to the results. For best performance with frequently used filters, consider using the label-based approach described above.

Tuning

The StreamingDiskANN index comes with smart defaults but also the ability to customize its behavior. There are two types of parameters: index build-time parameters that are specified when an index is created and query-time parameters that can be tuned when querying an index.

We suggest setting the index build-time paramers for major changes to index operations while query-time parameters can be used to tune the accuracy/performance tradeoff for individual queries.

We expect most people to tune the query-time parameters (if any) and leave the index build time parameters set to default.

StreamingDiskANN index build-time parameters

The StreamingDiskANN index build process can be memory-intensive. You may need to increase the maintenance_work_mem parameter to improve build performance. For example:

SET maintenance_work_mem = '2GB';

This parameter controls the maximum amount of memory to be used by maintenance operations, including index builds. The default value is typically 64MB, which may be too low for building StreamingDiskANN indexes on large datasets.

These parameters can be set when an index is created.

An example of how to set the num_neighbors parameter is:

CREATE INDEX document_embedding_idx ON document_embedding
USING diskann (embedding) WITH(num_neighbors=50);

An example of creating an index with label-based filtering:

CREATE INDEX document_embedding_idx ON document_embedding
USING diskann (embedding vector_cosine_ops, labels);

Parallel Index Build Parameters

pgvectorscale supports parallel index building to improve performance on large datasets. These parameters control parallel build behavior:

Parallel building is automatically enabled when:

• The feature is compiled in (enabled by default)
• The table has no labels (label-based filtering not yet supported in parallel)
• Using SBQ compression storage layout
• The number of vectors meets the min_vectors_for_parallel_build threshold

You can set these parameters using SET before creating an index:

-- Enable parallel building for smaller datasets
SET diskann.min_vectors_for_parallel_build = 10000;

-- Use 4 parallel workers regardless of PostgreSQL's automatic determination
SET diskann.force_parallel_workers = 4;

-- Adjust cache flushing frequency for memory usage (flush after processing 5% of vectors)
SET diskann.parallel_flush_interval = 0.05;

CREATE INDEX ON my_table USING diskann (embedding vector_cosine_ops);

StreamingDiskANN query-time parameters

You can also set two parameters to control the accuracy vs. query speed trade-off at query time. We suggest adjusting diskann.query_rescore to fine-tune accuracy.

You can set the value by using SET before executing a query. For example:

SET diskann.query_rescore = 400;

Note the SET command applies to the entire session (database connection) from the point of execution. You can use a transaction-local variant using LOCAL which will be reset after the end of the transaction:

BEGIN;
SET LOCAL diskann.query_search_list_size= 10;
SELECT * FROM document_embedding ORDER BY embedding <=> $1 LIMIT 10
COMMIT;

Null Value Handling

• Null vectors are not indexed
• Null labels are treated as empty arrays
• Null values in label arrays are ignored

ORDER BY vector distance

pgvectorscale's diskann index uses relaxed ordering which allows results to be slightly out of order by distance. This is analogous to using iterative scan with relaxed ordering with pgvector's ivfflat or hnsw indexes.

If you need strict ordering you can use a materialized CTE:

WITH relaxed_results AS MATERIALIZED (
    SELECT id, embedding <=> '[1,2,3]' AS distance
    FROM items
    WHERE category_id = 123
    ORDER BY distance
    LIMIT 5
) SELECT * FROM relaxed_results ORDER BY distance;

Index on an UNLOGGED table

Creating an index on an UNLOGGED table is currently not supported. Trying will yield the error:

ERROR:  ambuildempty: not yet implemented

Get involved

pgvectorscale is still at an early stage. Now is a great time to help shape the direction of this project; we are currently deciding priorities. Have a look at the list of features we're thinking of working on. Feel free to comment, expand the list, or hop on the Discussions forum.

About Timescale

Timescale is a PostgreSQL cloud company. To learn more visit the timescale.com.

Timescale Cloud is a high-performance, developer focused, cloud platform that provides PostgreSQL services for the most demanding AI, time-series, analytics, and event workloads. Timescale Cloud is ideal for production applications and provides high availability, streaming backups, upgrades over time, roles and permissions, and great security.