Assistant orchestrator implementation

Hi community,

We have a system that interacts with users using multiple contextual assistants. For example, one assistant answers nutrition questions and another handles training questions.

We need an orchestrator to decide which assistant should reply to the user. I thought about using an assistant as the orchestrator because it can see all new messages plus the thread and call the appropriate assistant.

Is this solution overkill (considering this orchestrator assistant will never reply directly)? Should I implement a simpler service that gets the thread, adds the message, and initiates the correct assistant (meaning we call the appropriate function)?



Why not just have one assistant answer both nutrition and training questions?

@ryan4ai my experience is that the bigger the assistant becomes the less acurate is, so we divided the porducto in different contextes.

Unless it’s mission critical that accuracy is perfect, I would build a single assistant that with future smarter models will eventually become what you want. I.e build it under the assumption in a year the Ai will easily handle both training and nutrition in a single assistant.

It is highly critical to be accured.

If you are concerned about token costs, then I think that is a valid solution. Many Assistant frameworks use similar patterns. Additionally, you can probably optimize what is sent to the orchestrator. Perhaps they don’t need everything, of most recent parts or just summaries. Good luck with your product.

1 Like