Lantern Extras

This extension makes it easy to experiment with embeddings from inside a Postgres database. We use this extension along with Lantern to make vector operations performant. But all the helpers here are standalone and may be used without the main database.

NOTE: Functions defined in this extension use Postgres in ways Postgres is usually not used. Some calls may result in large file downloads, or CPU-intensive model inference operations. Keep this in mind when using this extension a shared Postgres environment.

Features

• Streaming download of vector embeddings in archived and uncompressed formats
• Streaming download of various standard vector benchmark datasets
  • SIFT
  • GIST
• Generation of various various embeddings for data stored in Postgres tables without leaving the database

Examples

-- parse the first 41 vectors from the uncompressed .fvecs vector dataset on server machine
SELECT parse_fvecs('/tmp/rustftp/siftsmall/siftsmall_base.fvecs', 41);

-- load the first 10k vectors from the uncompressed vector dataset into a table named sift
SELECT * INTO sift FROM parse_fvecs('/tmp/rustftp/siftsmall/siftsmall_base.fvecs', 10000);

-- load SIFT dataset ground truth vectors into a table from an online ftp archive
SELECT query,
       true_nearest INTO sift_ground
FROM get_sift_groundtruth('ftp://host/path/to/siftsmall.tar.gz');

-- generate CLIP embeddings for columns of a postgres table
SELECT abstract,
       introduction,
       figure1,
       clip_text(abstract) AS abstract_ai,
       clip_text(introduction) AS introduction_ai,
       clip_image(figure1) AS figure1_ai
INTO papers_augmented
FROM papers;

-- generate embeddings from other models which can be extended

-- generate text embedding
SELECT text_embedding('BAAI/bge-base-en', 'My text input');
-- generate image embedding with image url
SELECT image_embedding('clip/ViT-B-32-visual', 'https://link-to-your-image');
-- generate image embedding with image path (this path should be accessible from postgres server)
SELECT image_embedding('clip/ViT-B-32-visual', '/path/to/image/in-postgres-server');
-- get available list of models
SELECT get_available_models();

Getting started

Installing from precompiled binaries

You can download precompiled binaries for Mac and linux from Github releases page. Make sure postgres is installed in your environment and pg_config is accessible form $PATH. Unzip the release archive from lantern_extras the directory run:

make install

Building from source

Click to expand

You should have onnxruntime in your system in order to run the extension. You can download the onnxruntime binary realease from GitHub https://github.com/microsoft/onnxruntime/releases/tag/v1.16.1 and place it somewhere in your system (e.g. /usr/lib/onnxruntime)

Then you should export these 2 environment variables

export ORT_STRATEGY=system
export ORT_DYLIB_PATH=/usr/local/lib/onnxruntime/lib/libonnxruntime.so

In some systems you will need to specify dlopen search path, so the extension could load ort inside postgres.

To do that create a file /etc/ld.so.conf.d/onnx.conf with content /usr/local/lib/onnxruntime/lib and run ldconfig

This extension is written in Rust so requires Rust toolchain. Make sure Rust toolchain is installed before continuing The extension also uses pgrx. If pgrx is not already installed, use the following commands to install it:

#install pgrx prerequisites
sudo apt install pkg-config libssl-dev zlib1g-dev libreadline-dev
sudo apt-get install clang

#install pgrx itself
cargo install --locked cargo-pgrx --version 0.11.3
cargo pgrx init

Then, you can run the extension under development with the following

cargo pgrx run --package lantern_extras # runs in a testing environment

To package the extension run

cargo pgrx package --package lantern_extras

Initializing with psql

Once the extension is installed, in a psql shell or in your favorite SQL environment run:

CREATE EXTENSION lantern_extras;

Adding new models

To add new textual or visual models for generating vector embeddings you can follow this steps:

1. Find the model onnx file or convert it using optimum-cli. Example optimum-cli export onnx --model BAAI/bge-base-en onnx/
2. Host the onnx model
3. Add model information in MODEL_INFO_MAP under lantern_extras/src/encoder.rs
4. Add new image/text processor based on model inputs (you can check existing processors they might match the model) and then add the match arm in process_text or process_image function in EncoderService so it will run corresponding processor for model.

After this your model should be callable from SQL like

SELECT text_embedding('your/model_name', 'Your text');

Lantern Index Builder

Description

This is a CLI application that creates an index for Lantern outside of Postgres which can later be imported into Postgres. This allows for faster index creation through parallelization.

How to use

Installation

Run cargo install --path lantern_cli to install the binary

Usage

Run lantern-cli create-index --help to show the cli options.

Usage: lantern-cli create-index --uri <URI> --table <TABLE> --column <COLUMN> -m <M> --efc <EFC> --ef <EF> -d <DIMS> --metric-kind <METRIC_KIND> --out <OUT> --import

Example

lantern-cli create-index -u "postgresql://localhost/test" -t "small_world" -c "vec" -m 16 --ef 64 --efc 128 -d 3 --metric-kind cos --out /tmp/index.usearch --import

Notes

The index should be created from the same database on which it will be loaded, so row tids will match later.

Lantern Embeddings

Description

This is a CLI application that generates vector embeddings from your postgres data.

How to use

Installation

Run cargo install --path lantern_cli to install the binary if you have clonned the source code or cargo install --git https://github.com/lanterndata/lantern_extras.git to install from git.

or build and use the docker image

# Run with CPU version
docker run -v models-volume:/models --rm --network host lanterndata/lantern-cli create-embeddings --model 'BAAI/bge-large-en' --uri 'postgresql://postgres@host.docker.internal:5432/postgres' --table "wiki" --column "content" --out-column "content_embedding" --batch-size 40 --data-path /models

# Run with GPU verion
nvidia-docker run -v models-volume:/models --rm --network host lanterndata/lantern-cli:gpu create-embeddings  --model 'BAAI/bge-large-en' --uri 'postgresql://postgres@host.docker.internal:5432/postgres' --table "wiki" --column "content" --out-column "content_embedding" --batch-size 40 --data-path /models

nvidia-container-runtime is required for GPU version to work. You can check the GPU load using nvtop command (apt install nvtop)

Usage

Run lantern-cli create-embeddings --help to show the cli options. Run lantern-cli show-models to show available models.

Text Embedding Example

1. Create table with text data

CREATE TABLE articles (id SERIAL, description TEXT, embedding REAL[]);
INSERT INTO articles SELECT generate_series(0,999), 'My description column!';

Currently it is requried for table to have id column, so it could map the embedding with row when exporting output.

2. Run embedding generation

lantern-cli create-embeddings  --model 'clip/ViT-B-32-textual'  --uri 'postgresql://postgres:postgres@localhost:5432/test' --table "articles" --column "description" --out-column "embedding" --schema "public"

The output database, table and column names can be specified via --out-table, --out-uri, --out-column arguments. Check help for more info.

or you can export to csv file

lantern-cli create-embeddings  --model 'clip/ViT-B-32-textual'  --uri 'postgresql://postgres:postgres@localhost:5432/test' --table "articles" --column "description" --out-column embedding --out-csv "embeddings.csv" --schema "public"

Image Embedding Example

1. Create table with image uris data

CREATE TABLE images (id SERIAL, url TEXT, embedding REAL[]);
INSERT INTO images (url) VALUES ('https://cdn.pixabay.com/photo/2014/11/30/14/11/cat-551554_1280.jpg'), ('https://cdn.pixabay.com/photo/2016/12/13/05/15/puppy-1903313_1280.jpg');

2. Run embedding generation

lantern-cli create-embeddings  --model 'clip/ViT-B-32-visual'  --uri 'postgresql://postgres:postgres@localhost:5432/test' --table "images" --column "url" --out-column "embedding" --schema "public" --visual

OpenAI and Cohere Embeddings

Lantern CLI also supports generating OpenAI and Cohere embeddings via API. For that you should specify --runtime and --runtime-params arguments

# OpenAI
lantern-cli create-embeddings  --model 'openai/text-embedding-ada-002' --uri 'postgresql://postgres:postgres@localhost:5432/test' --table "images" --column "url" --out-column "embedding" --schema "public" --runtime openai --runtime-params '{ "api_token": "sk-xxx-xxxx" }'

# Cohere
lantern-cli create-embeddings  --model 'openai/text-embedding-ada-002' --uri 'postgresql://postgres:postgres@localhost:5432/test' --table "images" --column "url" --out-column "embedding" --schema "public" --runtime cohere --runtime-params '{ "api_token": "xxx-xxxx" }'

|> To get available runtimes use bash lantern-cli show-runtimes

Index Autotune

Lantern CLI supports autotuning HNSW index parameters. To use the functionality run

lantern-cli autotune-index -u 'postgresql://postgres:postgres@localhost:5432/test' -t "sift1m" -c "v" --metric-kind l2sq --test-data-size 10000 --k 20

To get full list of arguments use bash lantern-cli autotune-index -h

Daemon Mode

Lantern CLI can be used in daemon mode to continousely listen to postgres table and generate embeddings, external indexes or autotune jobs.

 lantern-cli start-daemon --uri 'postgres://postgres@localhost:5432/postgres' --embedding-table embedding_jobs --autotune-table index_autotune_jobs --autotune-results-table index_parameter_experiment_results --external-index-table external_index_jobs --schema public --log-level debug

This will set up trigger on specified table (lantern_jobs) and when new row will be inserted it will start embedding generation based on row data. After that the triggers will be set up in target table, so it will generate embeddings continousely for that table. The jobs table should have the following structure

-- Embedding Jobs Table should have the following structure:
CREATE TABLE "public"."embedding_jobs" (
    "id" SERIAL PRIMARY KEY,
    "database_id" text NOT NULL,
    "db_connection" text NOT NULL,
    "schema" text NOT NULL,
    "table" text NOT NULL,
    "runtime" text NOT NULL,
    "runtime_params" jsonb,
    "src_column" text NOT NULL,
    "dst_column" text NOT NULL,
    "embedding_model" text NOT NULL,
    "created_at" timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
    "updated_at" timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
    "canceled_at" timestamp,
    "init_started_at" timestamp,
    "init_finished_at" timestamp,
    "init_failed_at" timestamp,
    "init_failure_reason" text,
    "init_progress" int2 DEFAULT 0
);
-- External Index Jobs Table should have the following structure:
CREATE TABLE "public"."external_index_jobs" (
    "id" SERIAL PRIMARY KEY,
    "database_id" text NOT NULL,
    "db_connection" text NOT NULL,
    "schema" text NOT NULL,
    "table" text NOT NULL,
    "column" text NOT NULL,
    "index" text,
    "operator" text NOT NULL,
    "efc" INT NOT NULL,
    "ef" INT NOT NULL,
    "m" INT NOT NULL,
    "created_at" timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
    "updated_at" timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
    "canceled_at" timestamp,
    "started_at" timestamp,
    "finished_at" timestamp,
    "failed_at" timestamp,
    "failure_reason" text,
    "progress" INT2 DEFAULT 0
);
-- Autotune Jobs Table should have the following structure:
CREATE TABLE "public"."index_autotune_jobs" (
    "id" SERIAL PRIMARY KEY,
    "database_id" text NOT NULL,
    "db_connection" text NOT NULL,
    "schema" text NOT NULL,
    "table" text NOT NULL,
    "column" text NOT NULL,
    "operator" text NOT NULL,
    "target_recall" DOUBLE PRECISION NOT NULL,
    "embedding_model" text NULL,
    "k" int NOT NULL,
    "n" int NOT NULL,
    "create_index" bool NOT NULL,
    "created_at" timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
    "updated_at" timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP,
    "canceled_at" timestamp,
    "started_at" timestamp,
    "progress" INT2 DEFAULT 0,
    "finished_at" timestamp,
    "failed_at" timestamp,
    "failure_reason" text
);

-- Autotune results table should have the following structure:
CREATE TABLE "public"."index_parameter_experiment_results" (
     id SERIAL PRIMARY KEY,
     experiment_id INT NOT NULL, -- reference to job.id
     ef INT NOT NULL,
     efc INT  NOT NULL,
     m INT  NOT NULL,
     recall DOUBLE PRECISION NOT NULL,
     latency DOUBLE PRECISION NOT NULL,
     build_time DOUBLE PRECISION NULL
);

Lantern PQ

Description

Use external product quantization to compress table vectors using kmeans clustering.

Usage

Run lantern-cli pq-table --help to show the cli options.

Job can be run both on local instance and also using GCP batch jobs to parallelize the workload over handreds of VMs to speed up clustering.

To run locally use:

lantern-cli pq-table --uri 'postgres://postgres@127.0.0.1:5432/postgres' --table sift10k --column v --clusters 256 --splits 32

The job will be run on current machine utilizing all available cores.

For big datasets over 1M it is convinient to run the job using GCP batch jobs.
Make sure to have GCP credentials set-up before running this command:

lantern-cli pq-table --uri 'postgres://postgres@127.0.0.1:5432/postgres' --table sift10k --column v --clusters 256 --splits 32 --run-on-gcp

If you prefer to orchestrate task on your own on premise servers you need to do the following 3 steps:

1. Run setup job. This will create necessary tables and add pqvec column on target table

lantern-cli pq-table --uri 'postgres://postgres@127.0.0.1:5432/postgres' --table sift10k --column v --clusters 256 --splits 32 --skip-codebook-creation --skip-vector-compression

2. Run clustering job. This will create codebook for the table and export to postgres table

lantern-cli pq-table --uri 'postgres://postgres@127.0.0.1:5432/postgres' --table sift10k --column v --clusters 256 --splits 32 --skip-table-setup --skip-vector-compression --parallel-task-count 10 --subvector-id 0

In this case this command should be run 32 times for each subvector in range [0-31] and --parallel-task-count means at most we will run 10 tasks in parallel. This is used to not exceed max connection limit on postgres.

3. Run compression job. This will compress vectors using the generated codebook and export results under pqvec column

lantern-cli pq-table --uri 'postgres://postgres@127.0.0.1:5432/postgres' --table sift10k --column v --clusters 256 --splits 32 --skip-table-setup --skip-codebook-creation --parallel-task-count 10 --total-task-count 10 --compression-task-id 0

In this case this command should be run 10 times for each part of codebook in range [0-9] and --parallel-task-count means at most we will run 10 tasks in parallel. This is used to not exceed max connection limit on postgres.

Table should have primary key, in order for this job to work. If primary key is different than id provide it using --pk argument