I believe that many users have experienced interruptions He Tuber  in conversations during conversations with AI products. So, on the product side, how can we design and interact well in interruption scenarios to meet the potential demands of users? In this article, the author has sorted out and shared it, let’s take a look.

1. User scenario

In the process of dialogue and interaction between users and generative AI products like ChatGPT, the dialogue may be interrupted for some reasons. At this time, what is a better solution for designing the data flow logic and product interaction?

There are two types of interruption scenarios. The main classification is based on whether the network is cut off and whether the data transmission and interaction process are terminated.

Interruption scenario ①: network terminal, refresh browser, close browser

In this scenario, the user actively or passively interrupts the network connection with the product;

The scope of this article is that the user in the conversation has successfully sent the question and then disconnected from the network.

Interruption scene ②: switch browser tabs, switch to view conversation records

In this scenario, the user only switched the screen display, but did not cut off the network and terminate the data transmission and interaction process;

There may be situations where you do not stay in the interactive interface for a long time.

2. Demand analysis

Based on the above user usage scenario, we need to further analyze the user's usage needs and interaction psychology in this scenario. We found that both scenarios actively or passively interrupted the user's normal conversation path, making the user's experience unsmooth, resulting in interruptions, losses, lags, etc. We need to use scenario-based and reasonable product interaction. Design to help users complete interactive actions and feel good.

sers’ requirements for the product may be as follows:

When an interruption occurs, I need to clearly perceive that the situation has changed;

There needs to be a connection between the state at the time of interruption and the state after recovery to allow users to maintain cognitive continuity;

The data that has been interacted cannot be lost, can be found quickly, and the stability of the product can be felt.

3. Product Goals

In the interaction of conversational AI products, this kind of scenario of actively or passively interrupting the user's normal conversation path is common, especially for component-based and PC-side interfaces.

Based on the analysis of scenarios and user demands, the product needs to solve this problem and achieve the following goals:

Sorting out data flow relationships by scenario;

Analyze data logic and clarify data presentation methods;

Scenario-based interaction process;

Design of user paths.

4. Data logic

1. Data transmission method of AI dialogue

When users need to use GPT products, they usually need to enter a piece of text. Once the user successfully sends "this text as a question" to the server, the question is submitted to the large model at once.

However, the way the large model returns data is streaming. The large model calculates the output step by step, and the server outputs it word by word to the front end, and the front end displays it to the user through the typewriter effect.

During this process, even if the user is chatting at the terminal, the data output of the large model will not be terminated.

As shown below:

2. How to save data for AI conversations

In a user dialogue interaction, after the large model completes all content output, the complete data of this round of dialogue is submitted and saved to the DB (database) at one time by the server.

The data in the business database can be used for front-end display and other purposes of products, and can also be used for back-end product analysis, etc.

3. Data echo method of AI dialogue

In different scenarios, data is echoed back to the user interface in different ways.

Normal interface dialogue: typewriter effect is displayed word by word;

After the conversation is interrupted: the front-end polls to obtain the complete data output by the AI. Once obtained, it will be displayed on the front-end all at once, without a typewriter effect;

Disconnect the network: You can view the complete data through the conversation record. Once viewed, it will be displayed on the front end at once, without a typewriter effect.