[FR]: templating syntax for message generation

[tal32123]

Proposal summary

I'd like to dynamically create messages/message parts. Currently I have a data structure with arrays of attachments which if there are attachments in that array I'd want to send to the llm, but in the way it currently works, I cannot. I attached a demo CSV file to better illustrate a data type/schema id like to use.

Motivation

My goal is to let someone who isn't a developer do the prompt engineering. We have dynamic data though as mentioned before. The current method in the playground does not allow for dynamic data.

demo.csv

[Nimrod007]

Hey @tal32123 , The idea is to create a dataset with nested columns ? in this example images json list ?

when you say dynamic data in the playground, are you referring to creating a dataset for the non technical person to use ? how will the actual list of images (or list of attachment) get resolve in runtime ?

[talbananaz]

Hi @Nimrod007 thanks for the quick response! :D So I want to make a dataset with a unified structure like in the photo I attached where all the photos of a certain type will be in the array for that type and from there the non techies can play around with different prompts for example if to send all the images there or one image or none, etc to see what gets the best results.

For example, I send in the photos array, and the outputs are not working, one because there are several elements, the second because there are no elements. But I want to be able to send in all elements in the array, or not send if there are none. So, I'd have to have different messages/message parts for each attachment or something in vercel.

I know braintrust does jinja and mustache, but they don't let you make message parts dynamically to my knowledge.

Anther example would be letss say I have a zoo animal llm and I have a json of the data but also I have images in different arrays per category so sea animals array, land animals array, etc. And I want to see if sending the images will be better than JSON only or reverse or both, etc.

Another thing I'm doing is trying to see if batching questions will give decent results or not but I assume I can work with llm as judge on this with my data structure (that's the status/issues you see in the dataset image).

[Nimrod007]

@talbananaz , thanks for the example. we will see how we can get this done