I spent few weeks last year building a text to sql tool using codex model to do something like this but for all kinds of data sources. We pivoted away to something else for various reasons.

But your approach is much better. Pandas is used a lot. Build a tool on top of pandas. This is awesome.

Hi, cool stuff! Which LLM is being used in the background? I may have missed that info in the readme. Thanks!

Thanks!

Right now this is running off of GPT-3 (`text-davinci-003`) and via a small code change can run on codex (`code-davinci-002`) but the quality only improves a little bit with that change.

That said, this is the first version to show that the interface is viable; we are currently working on training our own foundation model on a hybrid tokenization of data and word tokens. I hope to improve this same toolkit in the future with these new models of our own that we are training.

digging thru the code https://github.com/approximatelabs/sketch/blob/9d567ec161015...

this seems to be using their gpt3 frameowrk: https://github.com/approximatelabs/lambdaprompt

which uses text-davinci-003 by default https://github.com/approximatelabs/lambdaprompt/blob/main/la...

https://hal9.com is focused on building data apps with LLMs, would love to explore integrating and contributing to Sketch. If this sounds interesting I’m at javier at hal9.ai

I don't have any experience with pandas. Can this directly connect to a db and run queries there (video seems to load a csv file).

If you can already write SQL to return a data set then you can get that set to pandas with pyodbc.

This is very cool. A useful case for gpt. One question / concern: isn't a person's address considered PII? Is the system flexible enough to add pre-statements such as "treat an address as PII"?

Related question: is this done on my machine or do I end up sending possible pii to a cloud service for evaluation?

This is sending summary statistics to a cloud machine by default (for ease of immediate use. https://github.com/approximatelabs/sketch#sketch-currently-u...

You can run using your own OpenAI key by setting 2 environment variables: (1) SKETCH_USE_REMOTE_LAMBDAPROMPT=False (2) OPENAI_API_KEY=YOUR_API_KEY

To run entirely locally (using your own GPU and a model like Bloom) one would have to add a new prompt type to `lambdaprompt` (the package that this depends on), have a machine with enough GPU resources, and then add a slight modification to sketch.

Not sure if this is a business you're building out of this or an experiment. For real use for any of my customers, I would need to run this entirely locally.

I think it's really awesome though!

Curious what "enough GPU resources" looks like? Would a GeForce RTX 40 or 30 series card with 12-24GB of RAM be sufficient per user running locally on their machine?

If this is using OpenAI which it seems is what it is using, It is only sending column headers / column names. Not the data. If you are concerned about column names, you could also mask it on the way out and back in. If you are looking for an end to end database connect and query, please reach out to me.

Damn, this looks pretty useful. I was finding that github copilot was really good at reading a CSV file and writing all the imports from that into migrations for DB import, but this looks like it does these data transformations even more robustly.

Is there any plans on getting this to work outside of the python/pandas ecosystem or is it intrinsically tied to that environment?

This is very cool! I've literally today been noodling with ideas to use probabilistic data structures in LLMs.

And TIL you can embed mp4s in a GitHub readme. Is that new?