Configuring pipeline and policies

Hello Rasa friends,

We are trying to build a conversational agent which can answer simple company policy related questions. We have a large set of documents which we are trying to turn into query based conversations using Rasa nlu and core. However, we are finding it difficult to choose from a range of pipeline and policies options. Seems several tokenizers and featurisers are getting deprecated so not sure which one to go for.

Any guidance on this based on your experiences would be highly appreciated.

Thank you.

Hi @kumaramitsrivastava,

assuming the Rasa version you are using is 1.8.0 and your data is english, Iā€™d recommend to start with a simple ConveRT pipeline depending on your specific setup. The ConveRT pipeline achieved awesome results lately.

The EmbeddingIntentClassifier is also deprecated and now corresponds to the new DIETClassifier which unites features needed for IntentClassification and EntityExtraction - as configured.

If you are not sure about the outcome, Iā€™d suggest to simply start with the mentioned setup and do an evaluation and post the results here such that we can discuss them and maybe finetune the pipeline based on your needs.

Did that help?

Kind regards
Julian

1 Like

Thank you Julian,

Yes 1.8.0 and data is only English. Let me try and get back to you as suggested.