Training is the heavy process that trains the neuronal network which is "behind the scenes", so in fact the exact idea is to do training once and then store the result and read it when using - then training is not needed.

Also I think the current nlp.js do not provide ways for an "iterative training" which retrains the neuronal network after a new sentence is added or such - it would empty the neuronal network and train it again from scratch.

If you look at the time for training each epoch, it says 0ms.
After Epoch 36 stops, it should be totally trained.
Then is stopped, training is done, and now is waiting for the user to provide a question.

If you provide your code, we can investigate what connector are you using, and show you how to chat with the bot.

Hi @ethindp,

• logs stop printing
  as @jesus-seijas-sp's said, the logs of your training look fine. The numbers you see there are like an average error of the each training step. The default threshold for the neural network error is 0.00005, which means that once the loss goes below that threshold, that error is considered "acceptable" and training finishes. That's why you see no more logs.

• corpus file question
  it's up to you whether you want to feed your neural network with a corpus file or adding the documents programatically. Having a corpus file, though, will probably make it easier for you to maintain and evolve it. To se how to use a corpus file, just check the examples

• changing the corpus
  as @Apollon77 said, changing the corpus implies a retrain of the full corpus. This is because adding a new utterance would impact the relevance of words in existing utterances, so the weights need to be recalculated. For intance, if you have a single phrase in your corpus like "I want a car", adding another utterance like "I want a bike", would probably reduce the relevance of the word "want", and increase the relevance for "car" and "bike"... so, as existing weights need to be modified, a full retrain is currently needed

Hi all,
Thank you -- I figured out that it was a problem with my code (specifically not with the training part but how I was handling the returned data). Thank you for the extra knowledge; I'm new to NLP, so this is all new to me. Is there documentation that describes the returned response object when processing text? When I was playing with it earlier, I saw a "classification" object with the intent that I had added, even with variations that I hadn't added (which is a great thing). But I'm not sure if I should pay attention to the "classification" or "answers" array.

Hi @ethindp ,

it depends on your use case, generally you'll want to use the answers. Anyway, by having a look at the content of each one, you should be able to determine which one you need to use.

I closed this at seems a topic already solved.