Graphlogic.ai documentation

How to succesfully train your Agent

Testing the Agent script

After the

@Script⁠

is assembled and the

@Agent⁠

is trained (for information on how to successfully train the

@Agent⁠

see

Requirements for successful Agent training⁠

), it must be tested.

To do this, open the

@Debug Widget⁠

, where you can chat with the

@Agent⁠

, by switching the DEBUG switch.

⁠

Important: if an hour or more has passed since your last message to the
@Agent⁠
in the
@Debug Widget⁠
, the
@Dialog⁠
will be closed and the
@Agent⁠
will stop responding. You will need to refresh the page for the
@Agent⁠
to start responding again.

In order to test the

@Agent⁠

as effectively as possible, you must complete the following steps:

Make the most complete list of questions for the

@Agent⁠

. It is necessary to think about how the real

@Bot User⁠

will formulate questions, and try to provide for most of them. Also, the list of questions must include phrases that the

@Agent⁠

should not recognize and should send to

@fallback⁠

Run a list of questions through the

@Agent⁠

, check the recognition of

@Intent⁠

s and calculate the percentage of correct answers of the

@Agent⁠

. If a phrase is not recognized (ends up in

@fallback⁠

) or ends up in the wrong

@Intent⁠

where it should, write it down and indicate the name of the

@Intent⁠

where it should have ended up. After all the

@Intent⁠

s have been checked, you need to enter the written phrases into the corresponding

@Intent⁠

s and retrain the

@Agent⁠

. If irrelevant phrases that the

@Agent⁠

should not recognize end up not in the

@fallback⁠

, but in the

@Intent⁠

s, after completing regression testing, you should try to select a more optimal

@Confidence Threshold⁠

value. More details in the section

Testing and improving Agent recognition: selection of Confidence threshold⁠

After retraining the

@Agent⁠

, go through the list of

@Intent⁠

s again and check each one. Recalculate the percentage of correct answers. If the

@Intent⁠

s continue to get confused, check the

@Training Dataset⁠

. If the

@Training Phrase⁠

in the selection of the

@Training Dataset⁠

of different

@Intent⁠

s is too similar, you need to either make the

@Training Dataset⁠

of these

@Intent⁠

s more different from each other (remove similar wording, add more different phrases), or combine the confusing

@Intent⁠

s into one.

If the

@Vocabulary⁠

is involved in the

@Script⁠

, the work of the

@Vocabulary⁠

is checked as follows: you need to ask questions in order to get into a specific

@Script Branch⁠

— one of the branches with a reference — or the ‘true’

@Script Branch⁠

. If a phrase does not end up in the

@Script Branch⁠

it should, then go to the

@Vocabulary⁠

and check whether the word used in the phrase is in the

@Vocabulary⁠

dataset. If not, add it; if there is, check whether the same word is in other

@Entity⁠

After saving the changes to the

@Vocabulary⁠

, go through all the

@Script Branch⁠

es again and, if you do not fall into the desired

@Script Branch⁠

, add the missing words to the

@Vocabulary⁠

It is also necessary to check the operation of all functionality. Particular attention should be paid to slots with complex functionality —

@Regular Expression Slot⁠

@Memory⁠

etc.

Check the speed of transfer to the operator, if such a transfer is provided.

Check the operation of the

@External Request⁠

s and the speed of integration with external services, if provided.

Conduct a test in each involved

@Project Channel⁠

l: messengers, widget, etc.

Check spelling, punctuation and grammar.

Important: in order for the changes to take effect, do not forget to retrain your
@Agent⁠
using the Train button in BotBuilder.

Want to print your doc?
This is not the way.

Try clicking the ⋯ next to your doc name or using a keyboard shortcut (

CtrlP

) instead.