/postgres-gpt

Postgres-GPT combines pgvector and OpenAI to make a SQL-based knowledge repository from Markdown files that allows easy access to relevant information from multiple documents in a natural language format.

Primary LanguagePLpgSQLMIT LicenseMIT

Postgres-GPT

Postgres-GPT is a Postgres UDF project that seamlessly integrates Postgres-Vector and OpenAI to establish a SQL-based knowledge repository from Markdown files.

With its document embedding functionality, this project enables users to effortlessly access relevant information directly from multiple documents in a natural language format.

For instance:

-- Generate document embedding
select ai_embedding_doc('/opt/gpt/incubator-paimon/docs', 'paimon');

-- Ask questions to documents in natural language 
select ai_ask_doc('paimon', 'How to create a Paimon catalog on S3?');

Untitled

This project was inspired by AskBend and DataBend AI function projects, for which I am very grateful to Databend for their exploration in this area.


Quick Start

Installation

To run Postgres-GPT, which is an image that already includes all UDF functions and relevant plugins, Docker is the recommended method of running it.

docker run --name postgres-gpt -p 5432:5432 \
-e OPENAI_API_KEY=<your openai-api key> \
-e POSTGRES_PASSWORD=postgres \
-d yulinying/postgres-gpt

Alternatively, you can use the yulinying/postgres-gpt-playground image that includes pre-embedded sample document data.

Natural Language Document Search

Build embedding for automatic document indexing and use natural language queries to return results based on the closest vector cosine distance to a few paragraphs of document content, which are reorganized into natural language answers.

  1. Build document embedding, where doc_dir needs to be changed to a local path accessible by the postgres instance. This process takes a few minutes and needs to be executed only once unless the document changes occur.
select ai_embedding_doc('<doc_dir>', '<doc_name>');
  1. Query documents:
select ai_ask_doc('<doc_name>', 'your question');

Query Data in Natural Language

Explore data that exists in the database using natural language and automatically transform the intention into the corresponding SQL query using metadata from the database.

select ai_to_sql('<your prompt>');

About Postgres-GPT-Playground

yulinying/postgres-gpt-playground is a postgres-gpt image that has been pre-populated with sample data, including embedded documentation for the already constructed apache/incubator-paimon project, and a set of datasets for testing ai_to_sql function.

-- run container
docker run --name postgres-gpt-playground -p 5432:5432 \
-e OPENAI_API_KEY=<your openai-api key> \
-e POSTGRES_PASSWORD=postgres \
-d yulinying/postgres-gpt-playground

Let's explore these data:

select ai_to_sql('List detailed information for the top 5 companies with the highest event budgets in 2017');
select ai_ask_doc('paimon', 'What is the purpose of the paimon?');
select ai_ask_doc('paimon', 'How to create a Paimon catalog on S3?');

Function List

ai_embedding_doc

Split all markdowns in the specified directory into embedding vectors and automatically store them in postgres.

ai_embedding_doc(doc_dir varchar, doc_table varchar)

Parameters:

  • doc_dir: the local file directory that needs to be built, which postgres can directly access;
  • doc_table: the name of the table where the document embedding is built, which will be automatically created;

Example:

select ai_embedding_doc('/opt/gpt/incubator-paimon/docs', 'paimon');

ai_ask_doc

Search for the text closest to the question as a prompt from the vector results stored in doc_table and obtain an answer through openai.

ai_ask_doc(doc_table varchar, question varchar, max_token integer, mode varchar )

Parameters:

  • doc_table: the name of the table where the document embedding has been built;
  • question: the user's natural language question;
  • max_token: the maximum length of the answer text in tokens;
  • mode: the openai model used, which includes the following optional values:
    • v1: Use the text-davinci-003 model based on Completion API;
    • v2 (default): Use gpt-3.5-turbo model based on ChatCompletion API;

Example:

select ai_ask_doc('paimon', 'What is the purpose of the paimon?');

ai_to_sql

Generate the corresponding query SQL based on natural language, using metadata in the specified schema.

ai_to_sql(prompt varchar, schema varchar)

Parameters:

  • prompt: the data query requirement described in natural language
  • schema: target schema name.

Example:

select ai_to_sql('List detailed information for the top 5 companies with the highest event budgets in 2017');

-- result:
SELECT c.name, SUM(cam.monthly_budget) AS total_budget
FROM companies c
INNER JOIN campaigns cam ON c.id = cam.company_id
WHERE cam.updated_at >= '2017-01-01' AND cam.updated_at < '2018-01-01'
GROUP BY c.name
ORDER BY total_budget DESC
LIMIT 5;

Manual Installation

postgres-gpt is actually composed of a series of pl/python functions that can be easily installed on an existing postgres instance.

  1. Install the following postgres plugins:
    • pgvector
    • plpython3u
  2. Install the following python libraries:
    • markdown-it-py
    • openai[datalib]
pip install 'openai[datalib]' markdown-it-py
  1. Execute all sql scripts under the postgres_gpt directory on the corresponding postgres database.

Docker Build

docker build .