Embeddable AI

  • 1.  New Watson Assistant Limits?

    Posted Thu January 20, 2022 06:29 AM
    Edited by System Admin Tue May 20, 2025 06:58 PM
    Hi, WA Experts,

           A client faced a problem that exceeds the limit of Intent user examples - 25,000 in WA on Cloud.
           Checked the doc. of WA on Cloud and got confused on the limit of WA. As below:
           1. release notes of 27 November 2018, it seems the Intent user examples is 25,000 per Instance
              https://cloud.ibm.com/docs/watson-assistant?topic=watson-assistant-watson-assistant-release-notes
            
           2. The Intent limit under Creating intents, it show the Intent example limit is 25,000 per skill 
              https://cloud.ibm.com/docs/assistant?topic=assistant-intents#intents-limits

           Which one is true?

            Another question - is Skill keep in new Watson Assistant? It seems the answer is "No" and "Actions" is instead of "Skill" now.

            Also, would like to check if there is any solution for having Intent example over 25,000. Thanks! 

    ------------------------------
    KEVIN LIN
    ------------------------------
    #BuildwithWatsonApps
    #EmbeddableAI


  • 2.  RE: New Watson Assistant Limits?

    Posted Fri January 21, 2022 09:32 AM
    Your customer has 25000 user intent examples.  This is the limit of what the tool will hold.  That is TOO MANY EXAMPLES.  Here is why:
     
      A Watson Assistant skill is a big classifier, and it can be thought of as a giant, intelligent sorting algorithm.  It takes a user statement and figures out what their intent is.  In most cases, we suggest between 5 and 100 intents for any given chatbot.  More intents than that, and your intents begin to overlap.  In extreme cases, I have seen us have up to 500 intents for a single chatbot, and still work well.  Any larger, and your intents overlap and you end up with a mess.

     Now when training each intent, we suggest between 7 and 25 different training examples.  If you use more examples than that for a particular intent, you begin to run into overfitting your model.  Which means that your chatbot effectiveness goes down.  Usually, I have customers use no more than 12 training phrases per intent.
     
      So if we look at a worst-case scenario, we would have 500 intents, with 25 training phrases for each intent, and we would end up with 12500 user intent examples.  Only half of what your customer is using.
     
      Often customers will want a chatbot to handle questions and comments for multiple different knowledge domains.  In these cases, we recommend using cascading classifiers - multiple skills - and this article from Cari Jacobs is excellent in outlining how you can approach that - https://medium.com/ibm-data-ai/training-considerations-for-multiple-skills-orchestration-5d33ef7e7936.
     

    ------------------------------
    Daniel Toczala
    Community Leader and Customer Success Manager - Watson
    dtoczala@us.ibm.com
    ------------------------------

    Original Message