Tracker that gets sent to action server keeps on growing - Rasa Open Source - Rasa Community Forum

👋 Introducing the Rasa Playground

Tracker that gets sent to action server keeps on growing

You have selected 0 posts.

791 views
2 links

post by eugenelin89 on Jan 20, 2020

I was troubleshooting a mysterious error. The channel is Facebook Messenger. After I interact my bot for a while, the action server stops getting request from Rasa.

When I put Rasa under debug mode, I noticed the error

rasa.utils.endpoints.ClientResponseError: 413, Request Entity Too Large

Then I realized that the tracker that gets sent to the action server keeps on growing as I interact with rasa, and once it reaches a certain size the request gets an error at the HTTP level before it even gets processed by the action server.

I am new to Rasa and might be looking at the wrong place but I looked everywhere in the document but can’t zero in on a setting that limit the tracker… Has anyone encountered the same issue?

Thanks!

post by mloubser on Jan 21, 2020

You might want to look at the max_history setting in your policy ( Policies).

post by tyd on Jan 21, 2020

Hi @eugenelin89! I would also recommend checking out a new feature we just shipped called conversation sessions. You can now define the period of inactivity after which a new conversation session is triggered, which will allow you to keep the tracker size down.

post by anyone on Dec 23, 2021

Could you please provide a working example? With the following configuration and rasa 2.8.7 this doesn’t make any difference at all:

policies:
  - name: MemoizationPolicy
  - name: TEDPolicy
    max_history: 5
    epochs: 100
    constrain_similarities: true
    model_confidence: linear_norm
  - name: RulePolicy
    core_fallback_threshold: 0.5
    core_fallback_action_name: action_default_fallback

Thank you for all your help!

post by anyone on Dec 23, 2021

Thank you for your reply, this is a good feature if there is inactivity in between user utterances. However it would be great to have a solution that is working if the user is constantly talking to the assistant, as it is the case in my current use case.