This is minimal change to get dispatch actually working in the logs so it can be tested. After a few debug sessions with Lars we have more changes planned.
When in training the bot returns the action the dispatcher model takes which is (What model to dispatch to?)
When in logs the bot recursively delegates to the subModels until an action on one of the leaf model is reached and this is returned.
Usages depends on generation of model/actions using UI feature. It currently generates text actions with special encoding to indicate they are dispatch actions but should introduce new action type in future.
Technical Changes:
Detection of Dispatch action (Encoding of text actions)
Splitting out the MessageState tracking from BotState to allow independent CLMemory intances with different keys. We needed the BotState, BotMemory to be specific to the dispatch models of the conversation, but we needed the message state tracking for each of these models shared since all messages are for the same conversation.
This is minimal change to get dispatch actually working in the logs so it can be tested. After a few debug sessions with Lars we have more changes planned.
When in training the bot returns the action the dispatcher model takes which is (What model to dispatch to?) When in logs the bot recursively delegates to the subModels until an action on one of the leaf model is reached and this is returned.
Usages depends on generation of model/actions using UI feature. It currently generates text actions with special encoding to indicate they are dispatch actions but should introduce new action type in future.
Technical Changes:
More changes to come.