Graphlogic.ai documentation

Other

Dialog design best practice

Before creating an
`@Agent`⁠

1. The dialog designer (designer of the

@Agent⁠

‘s

@Script⁠

) and the tester/group of dialog design testers are defined. 2. The case is analyzed. 3. The

@Agent⁠

's goal, objectives and target audience are defined. 4. The main steps of the

@Script⁠

are outlined. 5. It is determined with whom / what the

@Agent⁠

will interact (

@Bot User⁠

, Operator, database, external services) and in which channels. 6. Integration channels work and support all conceived functionality. 7. The

@Agent⁠

@Script⁠

is designed in draw.io / miro / other program / on paper. 8. For NLU Slots, a training sample has been formed from examples of phrases with which the

@Bot User⁠

s will formulate requests to the

@Agent⁠

. 9. The

@Script Branch⁠

is logically distributed by intents (similar intents are combined into one, spreading branches through clarifying questions / regular expressions, etc.).

Creating an Agent

The involvement of the
`@Bot User`⁠
is ensured

1. Onboarding in the

@Agent⁠

: the purpose of the

@Agent⁠

is revealed in the welcome message. 2. The

@Agent⁠

has hints that ensure a painless passage of the

@Script⁠

. 3. The

@Agent⁠

directs the

@Bot User⁠

, calls for action (CTA) and provides him with a default action (if these are buttons). 4. There is a collection of feedback on the quality of the

@Agent⁠

's communication with the

@Bot User⁠

s ("Evaluate me, whether I helped you, what you liked, what not, etc.").

Text messages

1. The volume of the text of one

@Agent⁠

's message is ideally no more than 280 characters, the volume of information is reduced (divided into paragraphs / several messages / bullet-list), if there is nowhere to reduce, then the "read more" button. 2. The essence comes first. 3. The style of messages depends on the case and the target audience (if the

@Agent⁠

is informational, we use the informational style; if the

@Agent⁠

is entertaining, we add expressions, but at the same time we evaluate when exactly the

@Bot User⁠

expects the

@Agent⁠

to communicate in a friendly tone). 4. The message style is maintained throughout the

@Script⁠

. 5. Spelling and punctuation have been checked, especially quotation marks and dashes. 6. The

@Agent⁠

's answers are complete sentences, clear and detailed, do not begin with the words "yes", "no", etc. 7. The

@Agent⁠

's messages are gender neutral, unless otherwise required. 8. Links and images are correctly designed using tags/codes/ hyperlinks.

Other

1. Adapted Small Talk: The

@Agent⁠

reacts to emotions, several options for each Small Talk phrase – Random Responses. 2. Fallback is tailored to the

@Bot User⁠

and explains exactly what went wrong. 3. There are Random responses in the Fallback slot, and they explain exactly what happened and why, as well as suggest further action or its choice. For example: "I'm just a bot, so sometimes I can't recognize your questions. Please rephrase the question, or I can transfer you to the operator" and the buttons "try again" and "transfer to the operator". 4. The

@Agent⁠

is not looped, there are no slots that return the

@Bot User⁠

to the loop (for example, in a slot with buttons for some answer, there is a return to the previous text slot). 5. After each added branch, testing is carried out to identify errors in the

@Script⁠

. 6. Intents are primarily verbs, and nouns are entities. 7. Dictionaries or regular expressions are used to read entities (for example, the

@Bot User⁠

's message "my printer is not working" does not give an answer "what exactly are you having problems with?" and the buttons "technique", "pass", "software", and the essence of the problem is immediately determined using dictionaries or regularok). 8. In the Dialog

@Script⁠

, the

@Bot User⁠

has the opportunity to go back a step and, depending on the case, return to the start. 9. Before Slot Filling, the

@Agent⁠

informs that at the end the

@Bot User⁠

will have the opportunity to check and, if necessary, change the entered data. 10. The messages for Slot Filling show progress, for example: "1/5 Enter your phone number", "2/5 Enter your email address", etc. 11. Added minimal small talk (hi, thanks, so far, what can you do).

After creating the
`@Agent`⁠
⁠

1. The dialog designer forms a test sample that differs from the training one. With further changes to the NLU model, autotests will run the test sample through the NLU model. 2. The dialog designer tests the

@Agent⁠

on a test sample before passing the dialog design to the tester: most replicas are recognized, all the necessary variables are saved, all Slots work as they should (for example, Wait For Reaction is not missed anywhere), all integrations with external services work as they should). 3. The tester tests the

@Agent⁠

, focusing more on the case and less on the intentions and not focusing on the training sample. The tester conducts not only a test for the performance of the case, but also for usability. 4. The dialog designer analyzes the test data and, if necessary, trains the

@Agent⁠

and completes the

@Script⁠

. 5. If there were a lot of changes in retraining, the dialog design tester retests the

@Agent⁠

. 6. The

@Agent⁠

, if necessary, is tested by other team members; 7. The

@Agent⁠

is tested exactly in the channel(s) in which it will be used.

Before creating an Agent

Creating an Agent