pg_vectorize: a VectorDB for Postgres
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.
pg_vectorize powers the VectorDB Stack on Tembo Cloud and is available in all hobby tier instances.
API Documentation: https://tembo.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:
Then connect to Postgres:
Enable the extension and its dependencies
Install into an existing Postgres instance
If you're installing in an existing Postgres instance, you will need the following dependencies: Rust: - [pgrx toolchain](https://github.com/pgcentralfoundation/pgrx) Postgres Extensions: - [pg_cron](https://github.com/citusdata/pg_cron) ^1.5 - [pgmq](https://github.com/tembo-io/pgmq) ^1 - [pgvector](https://github.com/pgvector/pgvector) ^0.5.0 Then set the following either in postgresql.conf or as a configuration parameter: 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.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;
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/all-MiniLM-L6-v2',
schedule => 'realtime'
);
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.
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;
Initialize the RAG project.
We'll use the sentence-transformers/all-MiniLM-L6-v2
model to generate embeddings on our source documents.
SELECT vectorize.init_rag(
agent_name => 'product_chat',
table_name => 'products',
"column" => 'context',
unique_record_id => 'product_id',
transformer => 'sentence-transformers/all-MiniLM-L6-v2'
);
Now we can ask questions of the products
table and get responses from the product_chat
agent using the openai/gpt-3.5-turbo
generative model.
SELECT vectorize.rag(
agent_name => 'product_chat',
query => 'What is a pencil?',
chat_model => 'openai/gpt-3.5-turbo'
) -> 'chat_response';
And to use a locally hosted Ollama service, change the chat_model
parameter:
SELECT vectorize.rag(
agent_name => 'product_chat',
query => 'What is a pencil?',
chat_model => 'ollama/wizardlm2:7b'
) -> 'chat_response';
" A pencil is a writing instrument that consists of a solid or gelignola wood core, known as the \"lead,\" encased in a cylindrical piece of breakable material (traditionally wood or plastic), which serves as the body of the pencil. The tip of the body is tapered to a point for writing, and it can mark paper with the imprint of the lead. When used on a sheet of paper, the combination of the pencil's lead and the paper creates a visible mark that is distinct from unmarked areas of the paper. Pencils are particularly well-suited for writing on paper, as they allow for precise control over the marks made."
:bulb: Note that the -> 'chat_response'
addition selects for that field of the JSON object output. Removing it will show the full JSON object, including information on which documents were included in the contextual prompt.
Updating Embeddings¶
When the source text data is updated, how and when the embeddings are updated is determined by the value set to the schedule
parameter in vectorize.table
and vectorize.init_rag
.
The default behavior is schedule => '* * * * *'
, which means the background worker process checks for changes every minute, and updates the embeddings accordingly. This method requires setting the updated_at_col
value to point to a colum on the table indicating the time that the input text columns were last changed. schedule
can be set to any cron-like value.
Alternatively, schedule => 'realtime
creates triggers on the source table and updates embeddings anytime new records are inserted to the source table or existing records are updated.
Statements below would will result in new embeddings being generated either immediately (schedule => 'realtime'
) or within the cron schedule set in the schedule
parameter.
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';
Directly Interact with LLMs¶
Sometimes you want more control over the handling of embeddings. For those situations you can directly call various LLM providers using SQL:
For text generation:
select vectorize.generate(
input => 'Tell me the difference between a cat and a dog in 1 sentence',
model => 'openai/gpt-4o'
);
generate
-----------------------------------------------------------------------------------------------------------
Cats are generally more independent and solitary, while dogs tend to be more social and loyal companions.
(1 row)
And for embedding generation: