Lantern is an open-source PostgreSQL database extension to store vector data, generate embeddings, and handle vector search operations.
It provides a new index type for vector columns called hnsw
which speeds up ORDER BY ... LIMIT
queries.
Lantern builds and uses usearch, a single-header state-of-the-art HNSW implementation.
If you donβt have PostgreSQL already, use Lantern with Docker to get started quickly:
docker run --rm -p 5432:5432 -e "POSTGRES_USER=$USER" -e "POSTGRES_PASSWORD=postgres" -v ./lantern_data:/var/lib/postgresql/data lanterndata/lantern:latest-pg15
Then, you can connect to the database via postgresql://$USER:postgres@localhost/postgres
.
To install Lantern from source on top of your existing PostgreSQL:
git clone --recursive https://github.com/lanterndata/lantern.git
cd lantern
mkdir build
cd build
cmake ..
make install
To install Lantern using homebrew
:
brew tap lanterndata/lantern
brew install lantern && lantern_install
You can also install Lantern on top of PostgreSQL from our precompiled binaries via a single make install
.
Alternatively, you can use Lantern in one click using Replit.
Lantern retains the standard PostgreSQL interface, so it is compatible with all of your favorite tools in the PostgreSQL ecosystem.
First, enable Lantern in SQL (e.g. via psql
shell)
CREATE EXTENSION lantern;
Note: After running the above, lantern extension is only available on the current postgres DATABASE (single postgres instance may have multiple such DATABASES). When connecting to a different DATABASE, make sure to run the above command for the new one as well. For example:
CREATE DATABASE newdb;
\c newdb
CREATE EXTENSION lantern;
Create a table with a vector column and add your data
CREATE TABLE small_world (id integer, vector real[3]);
INSERT INTO small_world (id, vector) VALUES (0, '{0,0,0}'), (1, '{0,0,1}');
Create an hnsw
index on the table
CREATE INDEX ON small_world USING hnsw (vector);
Customize hnsw
index parameters depending on your vector data, such as the distance function (e.g., dist_l2sq_ops
), index construction parameters, and index search parameters.
CREATE INDEX ON small_world USING hnsw (vector dist_l2sq_ops)
WITH (M=2, ef_construction=10, ef=4, dim=3);
Start querying data
SET enable_seqscan = false;
SELECT id, l2sq_dist(vector, ARRAY[0,0,0]) AS dist
FROM small_world ORDER BY vector <-> ARRAY[0,0,0] LIMIT 1;
Lantern supports several distance functions in the index and it has 2 modes for operators:
-
lantern.pgvector_compat=TRUE
(default) In this mode there are 3 operators available<->
(l2sq),<=>
(cosine),<+>
(hamming).Note that in this mode, you need to use right operator in order to trigger an index scan.
-
lantern.pgvector_compat=FALSE
In this mode you only need to specify the distance function used for a column at index creation time. Lantern will automatically infer the distance function to use for search so you always use<?>
operator in search queries.Note that in this mode, the operator
<?>
is intended exclusively for use with index lookups. If you expect to not use the index in a query, use the distance function directly (e.g.l2sq_dist(v1, v2)
)
To switch between modes set
lantern.pgvector_compat
variable toTRUE
orFALSE
.
There are four defined operator classes that can be employed during index creation:
dist_l2sq_ops
: Default for the typereal[]
dist_vec_l2sq_ops
: Default for the typevector
dist_cos_ops
: Applicable to the typereal[]
dist_vec_cos_ops
: Applicable to the typevector
dist_hamming_ops
: Applicable to the typeinteger[]
The M
, ef
, and ef_construction
parameters control the performance of the HNSW algorithm for your use case.
- In general, lower
M
andef_construction
speed up index creation at the cost of recall. - Lower
M
andef
improve search speed and result in fewer shared buffer hits at the cost of recall. Tuning these parameters will require experimentation for your specific use case.
- If you have previously cloned Lantern and would like to update run
git pull && git submodule update --recursive
- Embedding generation for popular use cases (CLIP model, Hugging Face models, custom model)
- Interoperability with pgvector's data type, so anyone using pgvector can switch to Lantern
- Parallel index creation via an external indexer
- Ability to generate the index graph outside of the database server
- Support for creating the index outside of the database and inside another instance allows you to create an index without interrupting database workflows.
- See all of our helper functions to better enable your workflows
Important takeaways:
- There's three key metrics we track.
CREATE INDEX
time,SELECT
throughput, andSELECT
latency. - We match or outperform pgvector and pg_embedding (Neon) on all of these metrics.
- We plan to continue to make performance improvements to ensure we are the best performing database.
- Cloud-hosted version of Lantern - Sign up here
- Hardware-accelerated distance metrics, tailored for your CPU, enabling faster queries
- Templates and guides for building applications for different industries
- More tools for generating embeddings (support for third party model APIβs, more local models)
- Support for version control and A/B test embeddings
- Autotuned index type that will choose appropriate creation parameters
- Support for 1 byte and 2 byte vector elements, and up to 8000 dimensional vectors (PR #19)
- Request a feature at [email protected]
- GitHub issues: report bugs or issues with Lantern
- Need support? Contact [email protected]. We are happy to troubleshoot issues and advise on how to use Lantern for your use case
- We welcome community contributions! Feel free to open an issue or a PR. If you contact [email protected], we can find an open issue or project that fits you