Update entity recognition when bot is running

Rasa Open Source
1

When a user ask to login to a platform, I want to add entities that are currently available in that platform to my entities. So when I ask my bot is this an asset that he knows that something is an asset, because I just downloaded all my assets names in a list and my bot can extract this entity because he found an exact match in my list. I first tried this with lookup tables, but they are lacking this functionality.

So my question is: Has anyone ideas or figured out how to update your bot while running?

2

no, you need to retrain nlu, every time you update a list of entities

3

So I am writing a bot, to support a data governance system. And every time they add a value I should retrain the bot?

4

it’s hard to say without seeing train data, but how else would your bot know new values

5

you can adjust the lookup table file when updating, so it would be nice that it checks the file even when running. But current behavior is that it creates a regex pattern, loads all the values in when training and then never watch at this file again (unless you train again). When a user logs in I can safely update this file for all users since this is domain knowledge for everyone.

What I would like to achieve is, when the entity matches the regex for entity extraction (that you generated during training), that it looks into this file instead of the copy of this file that it loaded into memory.

6

You can create custom component for it

7

thx, I will look into that

8

I think something is missing here. The generated regex is generated using this file. That is why it doesn’t need look at this file again. It doesn’t look at the loaded copy of the file either. It just adds extra feature to the word matched by the regex.

However, if you want to dynamically extend the list of values that an entity can have, you have a couple of options:

  1. You can run the training periodically. This way new values will be available after some time.
  2. You can have a custom component that just blindly matches words in your file with the extracted entity. Not sure how well this can scale though.
9

well to reduce downtime I was thinking of a scheduled off moment to retrain everything and to keep the file that it blindly matches only to the differences. So after training this fille will be reset.