Create dbt package generator using GenAI and PyAirbyte

[aaronsteers]

Summary

The goal with this application is to take a specific raw data schema for a source being run with PyAirbyte and to auto-generate a simple dbt project for that data.

This could be the foundation of a new type of integration opportunity for Airbyte users.

Definition of Done

These are not specifically related to GenAI, but are the foundation of the code-gen:

• Solution should be written in Python.
• Solution can be written as a new feature in PyAirbyte or as a standalone project. (Author's preference, but probably easier/faster as a new project that just calls PyAirbyte.)
• Solution should be able to generate basic dbt project scaffold, including a basic "profiles" yaml and "dbt_project.yml". (Okay if these are hard-coded or hand-written as generic boilerplate.)
• Solution should be able to generate a "sources" yaml file for one or more sources that are being extracted using PyAirbyte. This should describe the tables being used.
• Solution should be able to be executed with dbt run, as proof of the working solution

The GenAI "code gen" application portion of this project is:

• Solution should be able to generate a dbt model (a .sql file) performing some basic transforms on top of the source table(s) defined in the "sources" yaml.
  • For instance, if the raw data is a 'sales' table, the LLM may create an aggregate table.
  • The LLM may also create 'stage' tables that take the raw schema and map the raw schema to new column names with conformed naming convention and/or conformed data types.
• Solution should use an LLM to generate the SQL.
• Instructions to the LLM can be hard-coded to one particular source, but the LLM should be doing the work of generating the SQL.

In terms of documentation:

• A README.md will be required for this project.
• A walkthrough tutorial explaining usage is also required. The walkthrough can exist within the README.md or can be provided in any other format, such as blog.
• A demo video walkthrough is optional, but not required.

Suggestions (Per Author's Discretion)

These are some suggestions - but are not required:

• We suggest using a simple source like source-faker, source-coin-api, or similar.
• We suggest using DuckDB as a backend - since it is easy to replicate results locally, doesn't required a paid account, and has good SQL support.

Resources to Assist

• PyAirbyte can be used to gather json schema for each stream.
• (@aaronsteers will add more resources and info here shortly.)

[Hashcode-Ankit]

hi @aaronsteers can I have more information on this topic would love to contribute to it.

[siddhant3030]

can you give us the summary on this?

[marcosmarxm]

@Hashcode-Ankit it is yours

[Hashcode-Ankit]

Hi @marcosmarxm thanks for assigning it to me @aaronsteers can I expect some more details on it? I have set up Airbyte and built some connectors using Airbyte. Need more information about how to approach this so that I can break down it into even more simple chunks.....

[aaronsteers]

@Hashcode-Ankit - I've updated the above with a description. Admittedly, this is a large and ambitious project. Let us know if it is still interesting, and/or if you have any questions or changes you would like to propose.

[Hashcode-Ankit]

@aaronsteers Yes as I mentioned on Slack as well, I am very much excited about it, and it matches my previous work on dbt and dbt-clickhouse.

[Hashcode-Ankit]

Hi @aaronsteers @marcosmarxm as discussed on slack i have pushed the code for it : https://github.com/Hashcode-Ankit/pyairbyte-dbt Need some more guidance on task to complete the exact use case @aaronsteers let me know if we can connect.