Our agent structure and the NLU model are getting bigger. Are there any design limitations, or recommended best practices for NLU models in Cognigy?
For optimal performance and user experience in terms of training times, Snapshot size, and Flow execution we would recommend not using more than 100 Intents, and 5000 sentences, in total per Flow.
Depending on the number of sentences and rules applied, Agents with 500 or more Intents will still train relatively fast, within a few minutes. There is a limit of 2000 sentences per Intent.
Training time and Snapshot size increase linearly with the number of Intents. Note that scoring performance is independent of the number of Intents, so this should not be a concern.
To keep training times, as well as; human effort in editing and maintenance manageable, we recommend using Cognigy's modular capabilities, for example to handle sub-topics within separate flows by using ‘Go To’ or ‘Execute Flow’ nodes.