Code360 powered by Coding Ninjas X Code360 powered by Coding Ninjas X
Table of contents
Train a Smart Reply model & manage allowlists
Creation & Training of a new model
Model allowlist management
Creation of conversation dataset
Creation of knowledge base
Create a conversation profile
Creating and editing a conversation profile 
How to use the Agent Assist simulator
Testing a conversation profile's performance
Frequently Asked Questions
Describe VPC.
What is load balancing?
What does Object Versioning entail?
Last Updated: Mar 27, 2024

Agent Assist

Author Shivani Singh
0 upvote
Leveraging ChatGPT - GenAI as a Microsoft Data Expert
Prerita Agarwal
Data Specialist @
23 Jul, 2024 @ 01:30 PM


When your human representatives are speaking with a consumer, Agent Assist employs machine learning technology to offer suggestions. Because suggestions are predicated on your own provided data, they can be customized to meet your unique business requirements. You can customize the suggested content that is given to your agents using a variety of elements in Agent Assist to offer various types of suggestions. For instance, the Smart Reply function allows you to configure suggestions. When a user asks a question, Agent Assist can offer the human agent several possible answers.

An interaction among two or more individuals is represented by a Conversation. A message is generated and saved for the conversation every time a conversation participant's utterance is provided to Agent Assist as part of a Conversation resource. A conversation profile is used to construct each interaction. The criteria contained in each conversation profile govern the suggestions made to the human agent throughout a chat.

Each conversation has three participants: END_USER, AUTOMATED_AGENT, and HUMAN_AGENT. You can create and customize virtual agents called Dialogflow agents to have interactions with your end users.

Having a conversation with an end user involves several steps:

  • Stage of connection: The end user starts a chat conversation.
  • Stage of the virtual agent: The user is linked to a Dialogflow virtual agent. The virtual agent interacts with the user and tries to answer all questions and fulfill all requests.
  • Stage of handoff: If the virtual agent is unable to resolve all customer issues, it is then given to a human. 
  • Agent Assist stage: The final is linked to one or more human agents during the agent assist stage. Agent Assist provides the human agent with pertinent documents and response recommendations in real-time.
  • Termination stage: The end user is cut off from the text chat at this point.

Train a Smart Reply model & manage allowlists

We will go over how to train and deploy a model using the Agent Assist Console in this section of the blog.

Creation & Training of a new model

Step 1: Go to the Agent Assist console and select the Models menu item on the page's far left margin. All of your models are visible in the Models menu. Click the +Create new button in the upper right to start a new model.

Source: Agent Assist

Step 2: Choose Smart Reply under Model type.

Step 3: Fill out the Name area with a special name for your new model.

Step 4: Decide the name(s) to use for the chat dataset(s) you intend to train your model. A pop-up menu with a list of all the datasets you've already produced shows when you click on the training dataset field. You are free to choose as many datasets as you want.

Step 5: Press Create to start building your new model. Until the new model is trained, its status will first appear as Pending and then as Creating.

Click on the indicator of three vertical dots that correspond to your trained model at the top right of the page, then select Deploy to use it.

Model allowlist management

A list of potential ideas that could be surfaced to agents during a discussion is also generated when a Smart Reply model is constructed using the Agent Assist interface. The allowlist is the name of this resource. The allowlist includes each and every proposal that could be made based on the chat dataset you choose.

  • Go to the Models menu, then click on the model's name to open the model's allowlist. Each of the three categories in the allowlist—Unreviewed, Allowed, and Blocked—has its own tab. 
  • Agents cannot be given blocked suggestions during runtime. 
  • During a dialogue, only ideas that are included on the Allowed and Unreviewed lists are brought forward. 
  • When a model is generated, every recommendation on the allowlist is set to Unreviewed by default.
Get the tech career you deserve, faster!
Connect with our expert counsellors to understand how to hack your way to success
User rating 4.7/5
1:1 doubt support
95% placement record
Akash Pal
Senior Software Engineer
326% Hike After Job Bootcamp
Himanshu Gusain
Programmer Analyst
32 LPA After Job Bootcamp
After Job

Creation of conversation dataset

Transcripts of conversations are part of a conversation dataset. With the help of this data, a Smart Reply model is trained to suggest text replies to human chat agents when they are interacting with end users. In order to establish a discussion dataset to test your integration or observe how Smart Reply performs before uploading your own data, Agent Assist also offers publicly accessible conversation data.

Step 1: The following page displays when you create a new chat dataset:

Source: Agent Assist

Step 2: Give your new dataset a name and optional description. Enter the URI of the storage bucket containing your conversation transcripts in the Conversation data section. The usage of the * sign for wildcard matching is supported by Agent Assist. The URI should be formatted as follows:

gs://<bucket name>/<object name>

Step 3: Press Create. On the Data menu page's Conversation datasets tab, your new dataset is now visible in the dataset list.

Creation of knowledge base

Step 1: Go to the Agent Assist Console. Click on the Data menu item on the page's far left side after choosing your GCP project. All of your data is displayed under the Data menu. There are two tabs, one for knowledge bases and the other for conversation datasets:

Source: Agent Assist

Step 2: At the top of the knowledge bases page, click the +Create new button after selecting knowledge bases: Enter a name for the knowledge base and select a language from the pop-up menu

Step 3: A list of all the documents in this knowledge base will appear, which is now empty. Click +Create New to add a document.

Source: Agent Assist

Step 4: If you're submitting FAQ papers, pick FAQ under choose a knowledge type.

Step 5: Select the location of the document you're adding under choose a file source.

Step 6: To add the document, click Create.

Create a conversation profile

A conversation profile sets up a number of settings that regulate the recommendations presented to an agent. The suggestions that are presented during runtime are controlled by these settings. Each profile sets either a human agent or a virtual agent from Dialogflow for a chat.

Creating and editing a conversation profile 

Step 1: Open the Agent Assist console, choose your project from the list, and then click Conversation profiles on the left sidebar menu.

Step 2: In the top right corner of the page, click +Create new. The next page is displayed:

Source: Agent Assist

Step 3: In the Display name box in the pop-up menu, give your discussion profile a distinctive name.

Step 4: From the list of available possibilities, select one or more suggestion-type options.

Step 5: Decide whether to use inline retrieval or pub/sub retrieval.

Step 6: (Optional) Switch to a Dialogflow virtual agent or enable sentiment analysis.

Step 7: Press Create. Before the conversation profile is usable, it can take a while.

How to use the Agent Assist simulator

Before putting your model into practice, you can preview its performance using the simulator included in the Agent Assist Console. Any Agent Assist feature can be used with the simulator. We'll go over the steps needed to run the Agent Assist simulator in this section of the blog.

Testing a conversation profile's performance

Step 1: Go to the Agent Assist Console and select Conversation profiles from the menu on the left-hand side of the screen.

Step 2: A list of your conversation profiles is visible on the conversation profiles page. Click the Use simulator button after selecting the three vertical dots.

Step 3: The window that displays offers input fields for responses from the consumer and the agent. Text input can be used to test your model. An article suggestion discussion profile is being tested using the example below:

Source: Agent Assist

Step 4: For Smart Reply models, possible human agent responses are presented in bubbles underneath the Agent text field. The suggested responses are regularly updated as the debate goes on. The conversation profile and model parameters influence the outcomes.

Frequently Asked Questions

Describe VPC.

Virtual Private Cloud is referred to as VPC. This virtual network provides access to numerous resources, including the Google Kubernetes Engine clusters, and computes Engine's VM instances. The VPC provides a lot of flexibility for managing the connections between workloads on a global or regional scale.

What is load balancing?

In a cloud-based computing environment, load balancing is a mechanism for allocating workloads and computer resources in order to manage requests. Because the workload is properly controlled through resource allocation, it offers a high return on investment at lower costs.

What does Object Versioning entail?

For the purpose of restoring overwritten and deleted items, object versioning is used. Versioning things makes them secure when rewritten or deleted, but it raises storage costs. When object versioning is enabled in the GCP bucket, anytime the item is deleted or overwritten, a non-common version of the object is created.


To sum it up, in this agent assist blog, firstly, we discussed how to train a smart reply model and manage all the allowlists. Then we discussed the creation of a conversation dataset, knowledge base, and conversation profile, and lastly, we saw how to use the agent assist simulator. 

For more content, Refer to our guided paths on Coding Ninjas Studio to upskill yourself.

Do upvote our blogs if you find them helpful and engaging!

Happy Learning!

Live masterclass