Question regarding RASA NLU run time

Hi everyone. I’m running Rasa NLU in production on the following spec: GCP e2-standard-8 machine - From which RASA is allocated the following: 4 CPU, 26 GB RAM

Config:

pipeline:

  • name: SpacyNLP model: “en_core_web_lg” case_sensitive: False
  • name: SpacyTokenizer
  • name: RegexFeaturizer
  • name: LanguageModelFeaturizer model_name: “bert” model_weights: “rasa/LaBSE”
  • name: CountVectorsFeaturizer analyzer: “char_wb” min_ngram: 1 max_ngram: 4
  • name: DIETClassifier model_confidence: softmax epochs: 100 batch_strategy: balanced constrain_similarities: True
  • name: EntitySynonymMapper
  • name: SpacyEntityExtractor case_sensitive: False
  • name: EntitySynonymMapper
  • name: CRFEntityExtractor
  • name: “DucklingEntityExtractor”

For each RASA classification (Intent prediction + entity extraction) it takes in average and median ~0.24 second. Is that considered a reasonable running time? Can we speed up the process?

Thank you!

Rasa Pro supports tracing to see a detailed breakdown of component performance for both training and prediction. You could also look at using a lighter weight pipeline like the LogisticRegressionClassifier.