Assistant API best practice

Hi everyone,
i’m trying the new beta assistant-api and i’d like to know your opinion on best practice with this tool.

Use case: help our customers on specific task or usage of our management software. Our software is splitted in macro-area and we create an assistent for each area and it answers questions against our technical data

First approach: we have created an assistant for each section (or macro-area) and uploaded all the documents (here comes the 20 documents limit per assistant…we need to join some of them) of that specific context.

First impression: the result is quite good, even becouse all the completion is handled by openai (multi-turn, vector similaity, etc…) but the risk of allucination is high if the assistant has many documents. Sometimes the first answer is not the right one, or partial right and only after some turn in conversation it gives the right one.

So what’s the best practice: are we supposed to increase assistant granularity creating other Assistants also by micro-area (in this way the assistant has less docs and the answer could be more specific…?)

Thank’s in advance as usual :wink:

2 Likes

Hi,

The fact is there are no best practices yet, we are all finding out the best things to do as we explore the technology! It certainly seems like a good idea to have assistants that specialise on a single dataset at this point, but that may not be the case with a little refinement or other ideas thrown into the mix.

Keep us updated with your findings and perhaps others have explored this area and can comment.

2 Likes

It seems that the use of your Assistant is quite similar to my project. I am planning to create multiple Assistants corresponding to each specialty, accompanied by uploading specific documents related to that field. Then, support users with question and answer based on the knowledge the assistant has. I am concerned about the cost for this Q&A, it seems that the fee will be charged based on each execution run of each message in the thread.

1 Like

I think it’s a little more expensive than doing vector similarity by ourselves, because we can limit the chunks of documents passed as context, but it seems that instead Assistant uses all he can get.

Smaller agents. Each should do the smallest possible task. E.g. break your macros into micros. Try it with one of your “macros” that seems problematic as a POC…

1 Like

Yes, it’s exactly what i’m doing. I’ve setted up an Assistant with function calling that works as a dispatcher to guide the user towards the right Assistant. … Work in progress … :wink:

2 Likes