vectorize 0.12.1

This Release
vectorize 0.12.1
Date
Status
Stable
Latest Stable
vectorize 0.14.1 —
Other Releases
Abstract
The simplest way to do vector search on Postgres
Description
Vectorize automates the transformation and orchestration of text to embeddings, allowing you to do vector and semantic search on existing data with as little as two function calls.
Released By
tembo
License
PostgreSQL
Resources
Special Files
Tags

Extensions

vectorize 0.12.1
The simplest way to do vector search on Postgres

Documentation

openai_embeddings
Vector Search with OpenAI
CONTRIBUTING
Contributing to pg_vectorize
README
vector-serve
rag
RAG
utilities
Utilities
search
Vector Search
index
PG Vectorize API Overview
index
index
sentence_transformers
Sentence Transformers

README

A Postgres extension that automates the transformation and orchestration of text to embeddings and provides hooks into the most popular LLMs. This allows you to do vector search and build LLM applications on existing data with as little as two function calls.

This project relies heavily on the work by pgvector for vector similarity search, pgmq for orchestration in background workers, and SentenceTransformers.


Static Badge PGXN version OSSRank

API Documentation: https://tembo-io.github.io/pg_vectorize/

Source: https://github.com/tembo-io/pg_vectorize

Features

  • Workflows for both vector search and RAG
  • Integrations with OpenAI’s embeddings and chat-completion endpoints and a self-hosted container for running Hugging Face Sentence-Transformers
  • Automated creation of Postgres triggers to keep your embeddings up to date
  • High level API - one function to initialize embeddings transformations, and another function to search

Table of Contents

Installation

The fastest way to get started is by running the Tembo docker container and the vector server with docker compose:

docker compose up -d

Then connect to Postgres:

docker compose exec -it postgres psql

Enable the extension and its dependencies

CREATE EXTENSION vectorize CASCADE;

If you’re installing in an existing Postgres instance, you will need the following dependencies:

Rust:

Postgres Extensions:

Then set the following either in postgresql.conf or as a configuration parameter:

-- requires restart of Postgres
alter system set shared_preload_libraries = 'vectorize,pg_cron';
alter system set cron.database_name = 'postgres'

And if you’re running the vector-serve container, set the following url as a configuration parameter in Postgres. The host may need to change from localhost to something else depending on where you are running the container.

alter system set vectorize.embedding_service_url = 'http://localhost:3000/v1/embeddings'

SELECT pg_reload_conf();

Vector Search Example

Text-to-embedding transformation can be done with either Hugging Face’s Sentence-Transformers or OpenAI’s embeddings. The following examples use Hugging Face’s Sentence-Transformers. See the project documentation for OpenAI examples.

Follow the installation steps if you haven’t already.

Setup a products table. Copy from the example data provided by the extension.

CREATE TABLE products (LIKE vectorize.example_products INCLUDING ALL);
INSERT INTO products SELECT * FROM vectorize.example_products;
SELECT * FROM products limit 2;
 product_id | product_name |                      description                       |        last_updated_at        
------------+--------------+--------------------------------------------------------+-------------------------------
          1 | Pencil       | Utensil used for writing and often works best on paper | 2023-07-26 17:20:43.639351-05
          2 | Laptop Stand | Elevated platform for laptops, enhancing ergonomics    | 2023-07-26 17:20:43.639351-05

Create a job to vectorize the products table. We’ll specify the tables primary key (product_id) and the columns that we want to search (product_name and description).

SELECT vectorize.table(
    job_name => 'product_search_hf',
    "table" => 'products',
    primary_key => 'product_id',
    columns => ARRAY['product_name', 'description'],
    transformer => 'sentence-transformers/multi-qa-MiniLM-L6-dot-v1'
);

This adds a new column to your table, in our case it is named product_search_embeddings, then populates that data with the transformed embeddings from the product_name and description columns.

Then search,

SELECT * FROM vectorize.search(
    job_name => 'product_search_hf',
    query => 'accessories for mobile devices',
    return_columns => ARRAY['product_id', 'product_name'],
    num_results => 3
);

                                       search_results                                        
---------------------------------------------------------------------------------------------
 {"product_id": 13, "product_name": "Phone Charger", "similarity_score": 0.8147814132322894}
 {"product_id": 6, "product_name": "Backpack", "similarity_score": 0.7743061352550308}
 {"product_id": 11, "product_name": "Stylus Pen", "similarity_score": 0.7709902653575383}

RAG Example

Ask raw text questions of the example products dataset and get chat responses from an OpenAI LLM.

Follow the installation steps if you haven’t already.

Set the OpenAI API key, this is required to for use with OpenAI’s chat-completion models.

ALTER SYSTEM SET vectorize.openai_key TO '<your api key>';
SELECT pg_reload_conf();

Create an example table if it does not already exist.

CREATE TABLE products (LIKE vectorize.example_products INCLUDING ALL);
INSERT INTO products SELECT * FROM vectorize.example_products;

Initialize a table for RAG. We’ll use an open source Sentence Transformer to generate embeddings.

Create a new column that we want to use as the context. In this case, we’ll concatenate both product_name and description.

ALTER TABLE products
ADD COLUMN context TEXT GENERATED ALWAYS AS (product_name || ': ' || description) STORED;
SELECT vectorize.init_rag(
    agent_name => 'product_chat',
    table_name => 'products',
    "column" => 'context',
    unique_record_id => 'product_id',
    transformer => 'sentence-transformers/all-MiniLM-L12-v2'
);
SELECT vectorize.rag(
    agent_name => 'product_chat',
    query => 'What is a pencil?'
) -> 'chat_response';
"A pencil is an item that is commonly used for writing and is known to be most effective on paper."

Trigger based updates

When vectorize job is set up as realtime (the default behavior, via vectorize.table(..., schedule => 'realtime')), vectorize will create triggers on your table that will keep your embeddings up to date. When the text inputs are updated or if new rows are inserted, the triggers handle creating a background job that updates the embeddings. Since the transformation is executed in a background job and the transformer model is invoked in a separate container, there is minimal impact on the performance of the update or insert statement.

INSERT INTO products (product_id, product_name, description)
VALUES (12345, 'pizza', 'dish of Italian origin consisting of a flattened disk of bread');

UPDATE products
SET description = 'sling made of fabric, rope, or netting, suspended between two or more points, used for swinging, sleeping, or resting'
WHERE product_name = 'Hammock';

Try it on Tembo Cloud

Try it for yourself! Install with a single click on a Vector DB Stack (or any other instance) in Tembo Cloud today.