RASA model's prediction accuracy drops when multiple users request concurrent

Does anyone ever ran into this problem: when only single user is requesting, RASA works just fine, intent prediction and action prediction are accurate, but when multiple user request concurrent, the performance can drop suddenly – wrong intent prediction or wrong action prediction.

I’ve thought that maybe messages from different users are “mixed”,like signal cross-talk. But I can’t locate this bug, I think the root casue is either socket or rasa, most likely RASA. I used a front-end server, a rasa server and an action server separately.

Can anyone give some suggestions on how to locate the issue or, any ideas to solve it?

1 Like

@nihuizhidao as long as you use different sender_ids there should be no problem with intent/action predictions. What kind of setup are you using?