Previous article[AI Learning] Nanny-level tutorial! Teach you step by step how to let Kimi become your own WeChat public account customer service for free (with practical verification results), we learned how to use coze+KimiDoWeChat Official Accountcustomer service.

However, from classmates and public accountsAI Customer ServiceJudging from the conversations, the following problems or capabilities are present:

① Not sure who I am. Sometimes I have hallucinations and make up a self-introduction.

② Unable to generate image.

③ The knowledge recognition ability is still at the level of January 2024 and cannot be queried online.

④ Unable to tell the weather.

⑤ The response speed is slow, and there is no dynamic feedback during the response.

⑥ Unable to recognize voice.

For the above 6 problems, solutions to the first 4 problems were found:

① Add the character description of the official account AI customer service.

② Added a plugin for generating images.

③ Add search plug-in.

④ Add weather query plug-in.

There is no solution to the fifth and sixth problems yet, which depends on the capability update of the Coze platform and WeChat public account.

Next, let’s learn togetherHow to solve the above 4 problems to evolve WeChat public account AI customer service capabilities (connect to the Internet, know the weather, and generate pictures)Bar.

PS: The WeChat public account AI customer service here isRefers to the public account AI customer service Bot created using the coze platform, hereinafter referred to as “Public Account AI Customer Service”.

Collection and analysis of public account AI customer service issues

The following are screenshots of specific needs that could not be met during the conversation between students and the official account AI customer service.

① Problems or missing abilities: not knowing who they are, experiencing hallucinations, making up self-introductions, and not being able to generate pictures.

② Problems or missing capabilities: Unable to generate images directly.

③ Problems or lack of capabilities: No online query capability, knowledge and information are only updated until January 2024.

④ Problems or missing abilities: Unable to provide the latest weather information, weather knowledge is limited to 2021.

⑤ Problems or lack of ability: slow response speed, no dynamic feedback during the answer generation process, so people feel that they have to wait a long time and don’t know whether the answer will be given.

⑥ Problems or lack of ability: Unable to recognize speech.

What can I do to solve a problem?

The problems have been collected and preliminarily analyzed. Next, how do we solve the problems?

OpenbuttonPlatform, click [Personal Space] on the left menu.

Find the conversational robot "Official Account AI Customer Service" created previously.

Enter the official account AI customer service editing interface.

Enter the following content in the Personality and Response Logic window to describe the personality, skills, and restrictions of the official account AI customer service.

Click the [Optimize] button, a prompt will pop up saying "Optimize content", click [Use].

Click [Plugins] in the middle skill section to expand it. You will see that there are currently no plug-ins. Click [+] on the right.

Enter the Add Plug-in interface.

In view of the lack of capabilities previously sorted out, the following plug-ins need to be selected here:

1. Tongyi Wanxiang: the ability to generate pictures.
2. Bing Search: Internet search capabilities.
3. Moji Weather: Weather query capability.

Going back to the skill interface, you can see that the Tongyi Wanxiang, Bing Search, and Moji Weather plug-ins have been added successfully.

Let's go to the preview debugging interface and ask the questions that our classmates asked the official account AI customer service again to see if we can get the expected answers.

Test item: Who are you?

The answer was as expected.

Test item: Generate a picture of a cute kitten standing in a square, looking up at the stars

The Tongyi Qianwen plug-in was successfully called and ran to completion, but no picture was seen.

I continued to give feedback but nothing was generated. This was embarrassing.

Next, continue with the following test.

First, search on Bing to find the latest news. For example, yesterday's news about AI being popular in China, and the news about Shengshu Technology releasing a large video model.

You can see that after calling Bing search, the correct answer can be given.

However, the returned message contains the first half of the call instructions, which is not very user-friendly and affects the viewing experience of the main content.

Next, try the weather query function. The Moji Weather official website shows that the weather in Beijing today (April 28) is as follows.

As you can see, today's Beijing weather forecast information is accurate.

However, just like the answer information from Bing search, the first half of the text containing the call instructions is not very user-friendly and affects the viewing experience of the main content.

Let's go back to the character setting and reply logic interface and add instructions in the restriction section:

When calling a plugin, give a direct answer without explaining the query process.

But there is no change in the resulting performance.

Calling the Wenshengtu plug-inUnable to generate image in the dialog windowand calling the plugin alwaysshow explain call query processGuotou tried to modify the character settings and reply logic in many ways, and also tried to call the workflow implementation, but still couldn't solve the problem.

Half a day passed, and Guotou fell into deep thought...

Suddenly I realized

Considering that Kimi itself does not support the function of generating pictures, could it be because the moonshot language model itself does not understand how to generate pictures in the dialogue window?

Secondly, the adaptation to the button's plug-in workflow is not good enough, so various adverse reactions occur?

Just do it when you think of it!

Pot HeadReplaced Kimi with the Skylark language model, the search plug-in is still Bing Search, the weather query plug-in is still Moji Weather, and the literary graph model is still Tongyi Wanxiang.

For personality and reply logic, click [Optimize] to generate a new description based on system suggestions.

Preview and debug again, and everything becomes clear.Fully meet the requirements. Haha ^_^

Questions about who you are

About the problem of generating pictures

Questions about online enquiries

Click the link to jump to the correct one, the news is correct.

Regarding weather query issues.

The corresponding content of Moji Weather is as follows and the information is accurate.

After confirming that everything is correct, click [Publish] in the upper right corner.

Enter the publishing interface, develop a good habit of generating publishing records, and click [Generate].

See the content of this update:

Bot Store: The model was switched from Moonshot (128K) to the Skylark language model, and the Tongyi Wanxiang series of plug-ins were added.

WeChat Official Account (Subscription Account): The character settings, reply logic, models, and plug-ins have been updated to provide users with better services.

Click [Publish] in the upper right corner.

The publication has been successfully submitted, click [Finish].

So far, some user problems have been solved.The evolved capabilities of public account AI customer serviceSuccessfully published to WeChat subscription account.

Verify the effect

I found that there were no problems with self-introduction, weather query, and Internet query, but generating pictures would get stuck at the "Continue" stage and there would be no response.

PS: I asked the official staff of KouZi, and they said that the optimization will be completed within 2 weeks after the holiday. (We will see the effect in two weeks)

Learning Summary

Let’s review and summarize what we have learned:

Question 1:I don’t know who I am, so sometimes I just make up a self-introduction.

Solution:Add AI customer service to public accountsCharacter Description.

Question 2:Unable to generate image.

Solution:Add a plugin to generate images, such as Tongyi WanxiangVincent graph plugin.

Question 3:The knowledge recognition ability is still at the level before January 2024 and cannot be queried online.

Solution:Add search plugins, such as Bing SearchPlugins.

Problem 4: Unable to tell the weather.

Solution:Add weather query plug-in, such asInk WeatherPlugins.

Question 5: The response speed is slow and there is no dynamic feedback during the response.

Solution:None yet. Students who have solutions are welcome to leave a message or send a private message to Guotou.

Problem 6: Unable to recognize speech.

Solution:None yet. Students who have solutions are welcome to leave a message or send a private message to Guotou.

Finally, I would like to share my experience and reminders about the pitfalls of Guotou (April 28, 2024):

① Currently, the moonshot language model used by Kimi does not support the image generation function, so it is impossible to generate images.

② The moonshot language model used by Kimi has been officially connected to Button for less than a week, and its adaptation to the plug-in workflow of Button is not good enough, which may cause some adverse reactions.

③ To solve problems ① and ②, directlySwitch the moonshot language model to the skylark language modelThat's it.

Now it's your turnBuild your own enhanced WeChat public account AI customer service (can connect to the Internet, know the weather, and generate pictures)La