Parameterize Style, End Goal, Agent Description

[ghiro994]

Hi, I haven’t been able to find a way to set default values for all users who automatically access Elysia for Style, End Goal, and Agent Description.

Is it possible to parameterize the default values of these three parameters via environment variables, or is there another recommended solution?

Pierluigi

[dannyjameswilliams]

Hey, thanks for your interest in Elysia!

Unfortunately, this is not possible in the current implementation, without every user loading the same global config.

This is a good idea though and I'll work on getting this implemented for the next version (using environment variables as defaults).

[ghiro994]

Hi again, thanks for your previous reply!

I’d also like to better understand how the three fields Style, End Goal, and Agent Description were originally conceived in terms of usage.

My intention is to make the best possible use of collections, and I’d like to know if these fields are also meant to:

describe the behavior or actions the LLM/agent should take,

or if they were rather designed as part of a workflow to give context to the model.

For example, when I query something like a customer record or related orders, I’d like to explain to the agent how to perform the query efficiently. Would these fields be the right place to define such guidance, or are they intended for something different?

Thanks a lot for any clarification, and sorry if this goes a bit beyond the scope of the original issue.

[dannyjameswilliams]

No problem, happy to help. Yes, so they all are intended as a different place to 'put instructions'. However, they are all added to the same prompt across all agents, to increase cohesiveness of the whole system.

• Style is just how the agent talks, essentially
• Agent description is more of an overall goal of the whole system. You could create a custom Elysia config with a custom set of tools designed to e.g. look up houseplants, with a description of when to use each tool (e.g. "identify plant with each user request, look up watering methods after identifying plants..." and so on). Essentially how to navigate its own custom decision tree.
• End goal is designed to be criteria on when the agent stops running, i.e. when the agent chooses that no more tools are required.

But like I said, they're just injected into each prompt - this could definitely be smarter and in a future version of Elysia, maybe these could be better handled.

For custom query instructions, I would recommend editing the metadata of your collections that it is relevant to. The LLM will generate a summary of the dataset, but you can edit this (in the app via the data tab -> metadata, and in Python via edit_preprocessed_collection). These will only be displayed when the collection is picked to be queried.