Skip to main content
Version: 2.0.0


This chapter will explain: policy definition and how to configure the agent's response policy on the platform

Session expiration time

In a real conversation, the user might not immediately respond to the robot's question, and the robot might just stop and wait like a human. If the user replies after a period of time, the session can continue. If the user does not respond, the bot automatically ends the session after a period of time.

To allow the bot to "pause" an intention to wait for some event to occur, the platform supports the session lifecycle mechanism. Cases where this mechanism can be used include:

  • The user does not respond in time after the robot asks.
  • Other intentions interrupt the current flow. The platform allows to be interrupted by another intention in the middle of the current dialogue flow running, thus running "another intention". Let's say the user inserts a question and answer into the flow.

Click "Build-Policies" in the agent menu, and you can see the session expiration time configuration item, and it can be configured for a time range of [1-1440] minutes expiration-time

In any of these cases, the current session is "paused". After that:

  • If the skill is returned during the lifetime of the session, the flow starts at the "suspended" position and continues backwards.
  • If the time has exceeded the session expiration time, then:
    • The skill process corresponding to the current intention will end;
    • Previously collected information like slots that was not integrated by subsequent processes is emptied;
    • The user must trigger the intent again to return back to the intent;
    • After refiring, the dialogue will be executed from the starting point of the flow.


When agent cannot give an accurate answer or thinks that the user's question does not need to be answered, it will give the user a default feedback by trigger a fallback action.

FallBack policy

  • Less than the minimum score: The intent that the bot calculates most similar to the user's query is usually called Top1 recall. If the score of the Top1 recall is still lower than the FallBack score, the agent will consider that it failed to match the correct intention and give an accurate response.

  • Match to the out-of-scope intent: the trainer generally maintain out-of-scope intent with queries that are often triggered by error but are not within the scope of the current agent. When Top1 recalls this out-of-scope intent, the agent will consider that the current query does not need to be answered.

FallBack Score

As long as Top1 recall score is lower than the FallBack score, agent will trigger a fallback action directly, and the higher the Top1 score is, the higher the accuracy will be.

Therefore, the higher the FallBack score is, the higher the response accuracy will be, but the recall rate will also be reduced. Conversely, the lower the FallBack score is, the accuracy of the response will decrease, but the recall rate will increase.

The configurable score range is [0.00-1.00], and it is generally recommended to set the bottom score between 0.6 and 0.7.

FallBack Action

In order to insure the agent respond to the user in any case, the agent needs to generate a fallback response even when it don't have anything to say.

The platform provides two ways to respond by sending a message or triggering a skill:

  1. Sending a Message

For example, a common bottom reply may be set to: "Sorry, I am not sure about this question. I am still learning, can you ask me questions in another way?"


  1. Trigger a skill

A common fallback skill might be "hand-off". The skill process is: "When fallback, decide whether it is time for customer service, and choose to transfer to manual or directly reply."



After completing or changing the configuration, please remember to click the save button to save the agent's policies.