A Primer on Big Model Application Development : Creating Agents Based on Coze

Hello, everyone, I'm Edison.

I recently took a course "Getting Started with AI Agent in Practice" to learn how to create an AI Agent on the Coze platform, and found it very useful for us personally (C-suite users), so I'll share it with you.

What's Coze?

Coze (Buckle) is a new generation of AI application development platform developed by ByteDance, using this AI application development platform, no matter whether you have a coding foundation or not, you can quickly build all kinds of AI Bot based on large language models, and you can also release the Bot to other channels. For an AI Agent, the most important capabilities are task planning, invocation tools, knowledge base and memory capabilities, and these capabilities in Coze you don't need to pay attention to, have been encapsulated and provided to you, for you is transparent.

As shown in the figure above, we can set up a persona for the AI Agent we want to do, then register it with the tools or workflows we want to call, and also register it with an internal knowledge base (documents/pictures/forms, etc.), and if we want the ability to memorize, we can even add a database for it to use, and then test it through the debugging module, and a "universe's strongest IDE" for AI Agent will be just like this. A "universe's strongest IDE" for AI Agent is just like that.
There are currently two versions of Coze:

(1) Basic Edition: for individual and enterprise developers to taste the experience, all functions are free to use, but there is a certain limited amount, after exceeding the amount can not be used again, you need to switch to the professional version to continue to use.

(2) Professional Edition: for professional developers with higher demands for stability and usage, supporting higher team space capacity and free knowledge base capacity, with paid functions to guarantee professional-grade SLAs, unlimited frequency and total amount of invocation requests, and fees calculated according to actual usage.

Here, I used the basic version, mainly to taste the freshness, made a few DEMO experience the effect, the model used is mainly beanbag Function Call model. In the future, we may mainly try FastGPT or Dify built within the enterprise, or Microsoft AutoGen.

Below, take a look at some of the key capabilities Coze offers, mainly through these DEMOs I made.

Powerful workflow configuration

I created a City Weather Assistant Bot through Coze, using the workflow capabilities provided by Coze, as shown below, which is a workflow to get the weather forecast and parse it:

As you can see, with a simple workflow, we can quickly invoke the capabilities of the big model and the plugin (InkWeather), which in traditional coding scenarios would require programmers to deal with individually, but are now 0-code purely configurable.

Based on this workflow, I then configure the cue word with the persona and reply logic to complete the creation of a Bot. It's worth mentioning that for your cue words, Coze provides an optimization feature to make your cue words optimized according to best practices, which is a really useful feature.

The final realization is shown below:

Powerful image stream configuration

I created a Bot for a product image background replacement helper through Coze, using another powerful skill provided by Coze: image streaming. This is also a workflow, but its uses processing nodes specialized for image processing, such as image generation, background replacement, image quality enhancement, and so on. These features should be quite helpful for those who are doing social media operations.

The final result is shown below: I gave it the original image and a description of the background I wanted to replace, and it gave me a pretty good background replacement.

Fast knowledge base application

Coze supports knowledge bases in different formats, such as text (e.g. txt, pdf, doc, etc.), table (e.g. xls, etc.) and photo (e.g. png, jpg, etc.).

For example, I created a MongoDB Knowledge Assistant Bot that imports some of the pdf documents for MongoDB's architecture course:

The final result is shown below:

As another example, I created a Product Drawing Lookup Assistant Bot that allows me to quickly find the corresponding product drawing based on the library of product drawings I imported.

The effect is shown below:

Again, let's say I'm the owner of an online classroom, and I import the course order form (Excel) into the knowledge base, and by setting up the persona and response logic, I can implement a quick query function:

Transparent memory capacity

Let's say I'm the owner of an online classroom, I could use Coze to create an online customer service that talks to customers and tries to direct users to leave their name and contact information, which would require a database-like memory capability.

Once configured in this way, once a customer leaves contact information in a conversation, our Bot automatically stores it in a pre-set database:

Other capacities

For customer service Bot, voice capability is very important, in Coze you can support voice calls, and there are multiple accents to choose from, personally I think this is a very convenient support capability.

Post to Subscription

Coze can support posting to multiple platforms, and the concept of Agent Store may actually come in the future. However, my current favorite is the ability to publish directly to WeChat subscriptions, so that when people reply to me, they don't just have cold auto-replies, but emotionally valuable replies, FOR ALL OF YOU!

wrap-up

This article briefly introduces the main features of Coze (button), an AI application development platform, and through some DEMOs that I have learned and practiced to understand some of the core concepts involved in AI Agent development such as workflow, image flow, memory capacity, knowledge base, etc. I believe that it will be helpful to you in the future practice of AI Agent development.

Recommended Learning

Wenyang Zhou, Getting Started with AI Agent in Practice