Awesome

pgai is a PostgreSQL extension that simplifies data storage and retrieval for Retrieval Augmented Generation (RAG), and other AI applications. In particular, it automates the creation and sync of embeddings for your data stored in PostgreSQL, simplifies semantic search, and allows you to call LLM models from SQL.

Docker

See the install via docker guide for docker compose files and detailed container instructions.

Timescale Cloud

Try pgai on cloud by creating a free trial account on Timescale Cloud.

Installing pgai into an existing PostgreSQL instance (Linux / MacOS)

See the install from source guide for instructions on how to install pgai from source.

Quick Start

This section will walk you through the steps to get started with pgai and Ollama using docker and show you the major features of pgai. We also have a quick start with OpenAI and a quick start with Voyage AI.

Setup

1. Download the docker compose file file.

   curl -O https://raw.githubusercontent.com/timescale/pgai/main/examples/docker_compose_pgai_ollama/docker-compose.yml
   

2. Start the docker compose file.

   docker compose up -d
   

   This will start Ollama and a PostgreSQL instance with the pgai extension installed.

3. Download the Ollama models. We'll use the all-minilm model for embeddings and the tinyllama model for reasoning.

   docker compose exec ollama ollama pull all-minilm
   docker compose exec ollama ollama pull tinyllama
   

Create a table and run a vectorizer

1. Connect to the database in your local developer environment The easiest way connect to the database is with the following command: docker compose exec -it db psql.

   Alternatively, you can connect to the database with the following connection string: postgres://postgres:postgres@localhost:5432/postgres.

2. Enable pgai on your database

   CREATE EXTENSION IF NOT EXISTS ai CASCADE;
   

3. Create a table with the data you want to embed from a huggingface dataset

   We'll create a table named wiki from a few rows of the english-language wikimedia/wikipedia dataset.

   SELECT ai.load_dataset('wikimedia/wikipedia', '20231101.en', table_name=>'wiki', batch_size=>5, max_batches=>1);
   

   Related documentation: load dataset from huggingface.

   This table will contain the following columns: id, url, title, text. We'll create a primary key on the id column:

   ALTER TABLE wiki ADD PRIMARY KEY (id);
   

4. Create a vectorizer for wiki

   We'll create a vectorizer that will automatically create embeddings for the text column in the wiki table.

   SELECT ai.create_vectorizer(
        'wiki'::regclass,
        destination => 'wiki_embeddings',
        embedding => ai.embedding_ollama('all-minilm', 384),
        chunking => ai.chunking_recursive_character_text_splitter('text')
   );
   

   Related documentation: vectorizer usage guide and vectorizer API reference.

5. Check the progress of the vectorizer embedding creation

   select * from ai.vectorizer_status;
   

   The output will look like this:

   id	source_table	target_table	view	pending_items
   1	public.wiki	public.wiki_embeddings_store	public.wiki_embeddings	10000

   All the embeddings have been created when the pending_items column is 0. This may take a few minutes as the model is running locally and not on a GPU.

6. Search the embeddings

   We'll search the embeddings for the concept of "properties of light"

   SELECT title, chunk
   FROM wiki_embeddings 
   ORDER BY embedding <=> ai.ollama_embed('all-minilm', 'properties of light')
   LIMIT 1;
   

   Note the ai.ollama_embed function is used to call the all-minilm model. This is part of pgai's model calling capabilities.

   The output will look like this:

   title	chunk
   Albedo	Water reflects light very differently from typical terrestrial materials. The reflectivity of a water surface is calculated using the Fresnel equations....

7. Modify your data and have the vectorizer automatically update the embeddings

   We'll modify the data in the wiki table and have the vectorizer automatically update the embeddings.

   INSERT INTO wiki (id, url, title, text) VALUES (11,'https://en.wikipedia.org/wiki/Light', 'Properties of light', 'Light is a form of electromagnetic radiation that can be detected by the human eye. It is a key component of the electromagnetic spectrum, which includes radio waves, microwaves, infrared, ultraviolet, and X-rays.');
   

   And now you don't need to do anything to update the embeddings. The vectorizer will automatically create the embeddings for the new row with any intervention from you. After a few seconds, you can run the search query again to see the new embedding.

Generate a summary of the article in the database

1. Generate a summary of the article in the database

   We'll generate a summary of the search results using the ai.ollama_generate function (this will take a few minutes).

   SELECT answer->>'response' as summary
   FROM ai.ollama_generate('tinyllama', 
   'Please summarize: '|| (SELECT text FROM wiki WHERE title='Albedo')) as answer;
   

   This is just one example of model calling capabilities. Model calling can be used for a variety of tasks, including classification, summarization, moderation, and other forms of data enrichment.

Features

Working with embeddings generated from your data:

• Automatically create and sync vector embeddings for your data (learn more)
• Search your data using vector and semantic search (learn more)
• Implement Retrieval Augmented Generation inside a single SQL statement (learn more)
• Perform high-performance, cost-efficient ANN search on large vector workloads with pgvectorscale, which complements pgvector.

Leverage LLMs for data processing tasks:

• Retrieve LLM chat completions from models like Claude Sonnet 3.5, OpenAI GPT4o, Cohere Command, and Llama 3 (via Ollama). (learn more)
• Reason over your data and facilitate use cases like classification, summarization, and data enrichment on your existing relational data in PostgreSQL (see an example).

Useful utilities:

• Load datasets from Hugging Face into your database with ai.load_dataset.

Resources

Why we built it:

• Vector Databases Are the Wrong Abstraction
• pgai: Giving PostgreSQL Developers AI Engineering Superpowers

Quick start guides:

• The quick start with Ollama guide above
• Quick start with OpenAI
• Quick start with VoyageAI

Tutorials about pgai vectorizer:

• How to Automatically Create & Update Embeddings in PostgreSQL—With One SQL Query
• [video] Auto Create and Sync Vector Embeddings in 1 Line of SQL
• Which OpenAI Embedding Model Is Best for Your RAG App With Pgvector?
• Which RAG Chunking and Formatting Strategy Is Best for Your App With Pgvector
• Parsing All the Data With Open-Source Tools: Unstructured and Pgai

Tutorials about pgai model calling:

• In-Database AI Agents: Teaching Claude to Use Tools With Pgai
• Build Search and RAG Systems on PostgreSQL Using Cohere and Pgai
• Use Open-Source LLMs in PostgreSQL With Ollama and Pgai

Contributing: We welcome contributions to pgai! See the Contributing page for more information.

Automated embedding and semantic search

• Automatically create and sync vector embeddings for your data
• Search your data using vector and semantic search
• Implement Retrieval Augmented Generation inside a single SQL statement

Automatically create and sync LLM embeddings for your data

The pgvector and pgvectorscale extensions allow you to store vector embeddings in your database and perform fast and efficient vector search. The pgai Vectorizer builds on top of these extensions to automatically create and synchronize embeddings for any text data in your database.

With one line of code, you can define a vectorizer that creates embeddings for data in a table:

SELECT ai.create_vectorizer(
    <table_name>::regclass,
    destination => <embedding_table_name>,
    embedding => ai.embedding_ollama(<model_name>, <dimensions>),
    chunking => ai.chunking_recursive_character_text_splitter(<column_name>)
);

This newly created vectorizer will automatically track any changes to the data in the source table and update the destination embedding table with the new embeddings asynchronously.

Automate AI embedding with pgai Vectorizer shows you how to implement embeddings in your own data. On a self-hosted Postgres installation, you use a Vectorizer Worker to asynchronously processes your vectorizers. When you create Vectorizers in a Timescale Cloud database, embeddings are automatically created and synchronized in the background.

Note: Timescale Cloud currently supports embedding natively with OpenAI. To use Ollama on the data in your Timescale Cloud service, set scheduling => ai.scheduling_none() in the configuration for your service, then install the vectorizer worker locally and configure it to connect to your Timescale Cloud service.

Search your data using vector and semantic search

pgai exposes a set of functions to directly interact with the LLM models through SQL, enabling you to do semantic search directly in your database:

SELECT
   chunk,
   embedding <=> ai.ollama_embed(<embedding_model>, 'some-query') as distance
FROM <embedding_table>
ORDER BY distance
LIMIT 5;

This is a perfectly normal SQL query. You can combine it with where clauses and other SQL features to further refine your search. pgai solves the missing where clause in vector search problem for real.

Implement Retrieval Augmented Generation inside a single SQL statement

Similar to semantic search, pgai LLM functions enable you to implement RAG directly in your database. For example:

1. Create a RAG function:

   CREATE OR REPLACE FUNCTION generate_rag_response(query_text TEXT)
   RETURNS TEXT AS $$
   DECLARE
      context_chunks TEXT;
      response TEXT;
   BEGIN
      -- Perform similarity search to find relevant blog posts
      SELECT string_agg(title || ': ' || chunk, E'\n') INTO context_chunks
      FROM
      (
          SELECT title, chunk
          FROM blogs_embedding
          ORDER BY embedding <=> ai.ollama_embed('nomic-embed-text', query_text)
          LIMIT 3
      ) AS relevant_posts;
   
      -- Generate a summary using llama3
      SELECT ai.ollama_chat_complete
      ( 'llama3'
      , jsonb_build_array
        ( jsonb_build_object('role', 'system', 'content', 'you are a helpful assistant')
        , jsonb_build_object
          ('role', 'user'
          , 'content', query_text || E'\nUse the following context to respond.\n' || context_chunks
          )
        )
      )->'message'->>'content' INTO response;
   
      RETURN response;
   END;
   $$ LANGUAGE plpgsql;
   

2. Execute your function in a SQL query:

   SELECT generate_rag_response('Give me some startup advice');
   

Model calling

Model calling is a feature of pgai that allows you to call LLM models from SQL. This lets you leverage the power of LLMs for a variety of tasks, including classification, summarization, moderation, and other forms of data enrichment.

The following models are supported (click on the model to learn more):

Some examples:

• Learn how to moderate content directly in the database using triggers and background jobs.
• load datasets directly from Hugging Face into your database.
• Leverage LLMs for data processing tasks such as classification, summarization, and data enrichment (see the OpenAI example).

Get involved

pgai 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.

To get started, take a look at how to contribute and how to set up a dev/test environment.

About Timescale

Timescale is a PostgreSQL database 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.