Coze platform architecture, typical application cases, advantages and application scenarios

Coze platform architecture

Two platform forms

Coze official website version (cloud version)

Coze’s official website is an officially hosted online service. Users access the platform through a browser and directly use the officially provided models, application capabilities, online operating environment and related services, without having to install programs or maintain servers themselves.

The main features of the official website version include:

You can use it online after registering;
The platform environment is uniformly maintained by the government;
Suitable for rapid experience, course teaching and prototype verification;
You can directly use the functions and ecological resources provided by the official website;
No need to handle system upgrades, infrastructure and daily operations and maintenance yourself;
The actual available functions, resource quotas and publishing capabilities are subject to the current version of the official website.

Coze’s official website currently describes itself as an AI product for workplace tasks, and provides AI applications, work product generation, long-term tasks and related development capabilities.

Coze Studio (open source version)

Coze Studio is an open source version of the core engine of the Coze platform. Developers can deploy it to local computers, laboratory servers or corporate intranets to build an autonomous and controllable AI Agent development platform.

The main features of the open source version include:

Support privatization deployment;
Data and operating environment can be managed by the deployment party;
Can access your own models and business services;
Support functional modifications based on source code;
Suitable for teaching research, corporate intranet and secondary development;
Certain deployment, configuration and operation and maintenance capabilities are required.

Coze Studio officially stated that it provides core AI agent development technologies such as Prompt, RAG, Plugin and Workflow, and uses visual methods to support the development, debugging and deployment of agents, applications and workflows.

Comparison between official website version and open source version

Comparison Dimensions	Coze Official Website Edition (Cloud Edition)	Coze Studio (Open Source Edition)
use	Browser use online	Local or server deployment
environmental maintenance	Official unified maintenance	User maintenance
difficulty of getting started	Lower, you can use it after registration	Higher, requiring deployment and configuration
data storage	Processed by the cloud platform according to service rules	Can be managed by the deployer in its own environment
function acquisition	Directly use the functions currently provided by the official website	Use open source features and can be extended on your own
model configuration	Subject to the official website support scope and platform rules	Supported models can be configured based on deployment environment
Plug-ins and Services	Use the official website ecosystem and online services	Need to configure open source plug-ins or access your own API
secondary development	Mainly through platform configuration and open interfaces	Source code can be modified and deeply customized
operation and maintenance costs	lower	higher
applicable objects	Beginners, individual users, quick verification team	Technical teams, laboratories, corporate privatization projects
main advantage	Out of the box and easy to use	Self-controllable and easy to customize
main limiting	Affected by platform functions, quotas and service rules	Need to solve deployment, resource and operation and maintenance issues yourself

The two forms have similar AI Agent development concepts, but there may be differences in specific functions, official ecology, resource quotas, release channels and operation and maintenance methods. The latest official instructions of the corresponding version shall prevail when used.

Platform Capability Architecture

From the perspective of the AI application construction process, Coze can be divided into five collaborative capability layers.

flowchart TB
    A[用户与业务系统]

    B[应用与交互层<br/>智能体 Agent / AI 应用 App / 对话界面 / API]
    C[智能与编排层<br/>大语言模型 / 提示词 / 工作流]
    D[能力扩展层<br/>插件 / 工具 / 第三方 API]
    E[知识与数据层<br/>知识库 / 数据库 / 长期记忆]
    F[开发与发布支撑<br/>可视化工作台 / 调试 / 发布 / 管理]

    A --> B
    B --> C
    C --> D
    C --> E
    F -.支撑.-> B
    F -.支撑.-> C
    F -.支撑.-> D
    F -.支撑.-> E

Application and Interaction Layer

The application and interaction layer is directly oriented to end users and is responsible for undertaking user needs and displaying AI processing results.

This layer mainly includes:

Agent: Using natural language dialogue as the main interaction method, it can understand tasks, select tools, and return results;
AI application apps: built around specific business requirements and usually have clear input, output and visualization interfaces;
Dialogue interface: The entrance for questions and answers and tasks between users and agents;
API and SDK: Used to connect Coze’s agents or workflows to websites, management systems and other business software.

Agents are more suitable for open dialogues and dynamic tasks, and AI applications are more suitable for functional scenarios with clear business goals and relatively fixed processes.

Intelligence and orchestration layer

The intelligence and orchestration layer is the core of the platform and is responsible for understanding user intentions, organizing business logic, and generating end results.

This layer mainly includes:

Big language model;
Prompt words;
Agent decision logic;
Workflow.

The large language model is responsible for understanding, reasoning, and generation; prompt words define the role and behavior of the model; and the workflow is responsible for dividing complex tasks into multiple steps and executing them in turn according to determined logic.

Capability extension layer

The capability extension layer connects to the outside world through plug-ins and tools, expanding the large model from “only generating content” to “being able to obtain information and perform operations.”

For example:

Call search tools to get real-time news;
Call the weather service to check the weather;
Call map services to plan routes;
Call the image generation service to create pictures;
Call the enterprise API to query order or customer information;
Call mail, calendar and other services to complete office tasks.

Knowledge and data layer

The knowledge and data layer provides the agent with professional knowledge, business data and long-term status.

Mainly include:

Knowledge base: saves enterprise documents, product manuals, course materials and industry reports, and provides search and enhanced generation capabilities for agents;
Database: Store structured business data;
Long-term memory: Recording user preferences, historical tasks, and interaction information so that the agent can continue to use this information in subsequent sessions.

Development and release support

Development and publishing capabilities run through the entire application building process, mainly including:

Visualization workbench;
Drag-and-drop workflow canvas;
Real-time preview and debugging;
Application configuration and resource management;
Agent and application release;
API and SDK integration.

This layer lowers the development threshold and allows non-professional developers to participate in the design and iteration of AI applications.

Core components of the Coze platform

Big Language Model

The big language model is the intelligence core of the Coze platform and is mainly responsible for:

Understand natural language;
Identify user intent;
Generating text content;
Carry out logical reasoning;
Determine whether to invoke the plug-in or workflow;
Integrate the results returned by the knowledge base and tools.

Model capabilities directly affect the agent’s understanding accuracy, response quality and tool call stability.

Agent

Agents are AI projects that can interact with users through natural language and invoke knowledge bases, plug-ins, or workflows based on task needs.

A complete agent usually consists of the following:

Role and identity;
System prompt words;
Big language model;
Dialogue context;
knowledge base;
Plug-ins and tools;
Workflow;
Databases and memory;
Opening remarks and recommendation questions.

Agents are not simple question-and-answer robots. It can select different abilities based on user goals and complete the closed loop of “comprehension-decision-execution-feedback” tasks.

AI Application App

AI applications are independent projects built around specific business requirements and have clear input, output and processing logic.

Typical AI applications include:

AI search;
Translation tools;
report generator;
Data format conversion tool;
Food recording tools;
Content review tools.

Compared with agents, the functional boundaries of AI applications are usually clearer and suitable for providing standardized services.

Prompt word

Prompt words are used to define the agent’s identity, task goals, work steps, reply style and behavioral boundaries, and are important components in controlling the agent’s performance.

A clearly structured reminder usually includes:

Role positioning: Who is the agent;
Mission goal: What needs to be accomplished;
Workflow: What steps should be followed;
Available resources: What knowledge bases, plug-ins, and workflows can be called;
Output specification: language, format, length and structure requirements;
Behavior constraints: prohibited matters and exception handling methods.

Prompt words can constrain model behavior, but cannot replace program rights control, content review, and security verification.

Workflow Workflow

Workflow is used to split complex tasks into multiple executable nodes, and organize data flow and execution order through connections between nodes.

Common nodes include:

Large language model nodes;
Plug-in node;
Knowledge base retrieval node;
code node;
condition judgment node;
Circular node;
Data processing node;
Input and output nodes.

Workflow is suitable for performing tasks with clear processes, many steps, and high stability requirements, such as industry report generation, order processing, data analysis, and content review.

Plug-ins and Tools

A tool is the ability to complete a specific operation, such as querying the weather; a plug-in is a collection of capabilities composed of one or more related tools.

The main functions of the plug-in are:

Obtain real-time information beyond the scope of large model knowledge;
Call third-party services;
Connect with internal business systems;
Performs operations such as queries, calculations, builds, and notifications.

Official instructions point out that plug-in tools can extend the capabilities of large language models to connect to the outside world.

Knowledge base and RAG

The knowledge base is used to store internal information, course documents, product manuals, industry documents and web content.

When a user asks a question, the system can first retrieve relevant content from the knowledge base, and then submit the search results to the large language model to generate answers. This “retrieval first, then generation” approach belongs to RAG, that is, retrieval enhanced generation.

The main values of the knowledge base include:

Supplementing private knowledge of the enterprise that the model does not understand;
Improve the accuracy of answers in professional fields;
Reduce unfounded answers;
Let the agent answer questions based on the specified data;
Reduce the time cost for users to find large amounts of documents.

RAG can reduce model illusions, but it cannot completely eliminate errors. Important content still needs to be cited and manually reviewed.

Databases and long-term memory

Databases are used to hold structured information, and long-term memory is used to hold user preferences, historical interactions, and task status.

For example, a reading assistant can record:

title;
Read page numbers;
Reading time;
User notes;
Interest preferences.

When the user talks again, the agent can query historical data and continue to provide services, transforming from a one-time question and answer tool into a long-term personal assistant.

Debugging and release

Coze provides visual preview and debugging capabilities. Developers can test before release:

Whether the prompt words are clear;
Answer whether the style meets the requirements;
Whether the plug-in can be called normally;
Whether the workflow is performing as expected;
Whether the knowledge base search results are accurate;
Whether the abnormal situation has been handled reasonably.

After the test is completed, the agent or application can be released and used by users, or it can be connected to existing systems through APIs or SDK.

Core component collaboration relationships

core component	main role	typical collaboration object
Language Model	Understanding, Reasoning, Generating and Decision-making	Prompt Words, Agents, Workflow
agent	receives user tasks and organizes the execution of	models, knowledge bases, plug-ins, workflows, and memories
AI applies	to encapsulate specific business functions	models, workflows, APIs
prompt	specifies roles, tasks and output boundaries	model, agent
Workflow	Orchestrating multi-step business processes	Models, plug-ins, knowledge bases, codes
plug-ins and tools	Get real-time information or perform external operations	Agents, workflows, third-party systems
Knowledge Base	provides private knowledge and search basis	Agents, models
Database and Memory	saves business data and long-term status	Agents, users, applications
API and SDK	connects platform capabilities to other systems	websites, enterprise systems, and mobile applications

Typical application cases of Coze platform

Boasting about robots

business needs

Identify positive information in user expressions through dialogue and praise them in humorous, sincere and positive language.

use components

Agents;
Big language model;
System prompt words;
Opening remarks and recommendation questions;
Optional search plug-in.

achieve effects

The role, language style and answer boundaries of the “master of exaggeration” are set through prompts, so that the robot always provides positive feedback. This case is suitable for demonstrating basic capabilities such as agent creation, prompt writing, debugging and publishing.

intelligent customer service

business needs

Companies need to continuously handle a large number of repetitive inquiries, such as product functions, after-sales rules, order status and common faults.

use components

Agents;
Enterprise knowledge base;
Search or order query plug-ins;
Workflow;
API and SDK.

achieve effects

Smart customer service can answer common questions 24/7. When the question involves enterprise systems or product parameters, the answer is retrieved from the knowledge base; when it involves order status, the business system API is called; when complex problems are encountered, they can be forwarded to manual customer service through the workflow.

Enterprise Knowledge Question and Answer Assistant

business needs

Employees need to quickly find information from employee manuals, product documents, project materials, and rules and regulations.

use components

Agents;
knowledge base;
RAG;
Authority control;
Conversation interface.

achieve effects

Enterprises import internal documents into the knowledge base, and employees can directly ask “What are the travel expense reimbursement standards” or “What are the core parameters of Product A?” Agents retrieve relevant data and generate answers, reducing the cost of manual document search and repeated consultations.

AI news announcer

business needs

Users want to regularly get the latest news in specific areas and organize multiple information sources into concise summaries.

use components

Agents;
News search plug-in;
Workflow;
Big language model;
Timing tasks or external triggering capabilities.

achieve effects

The system calls the news search plug-in to obtain real-time information, filters, de-duplicates and classifies it through the workflow, and then generates the broadcast script from the big language model. This case demonstrates the combination of plug-ins, workflow and text generation capabilities.

Industry research report generation

business needs

Traditional industry research reports require multiple steps such as data search, data collation, outline design, content writing and format check, which takes a long time.

use components

Workflow;
Search plug-ins;
knowledge base;
Big language model;
Condition judgment and code nodes;
Document output tool.

achieve effects

The workflow divides report generation into:

Receive research topics;
Collect public information;
Searching internal knowledge bases;
Generate a report outline;
Write in chapters;
Check logic and format;
Output final report.

This case is suitable to demonstrate Coze’s ability to automate multi-step, complex knowledge work.

Intelligent sales system for automobile brands

business needs

Car sales personnel need to judge purchase intentions based on user browsing behavior, consultation content and model preferences, and provide personalized recommendations.

use components

Agents;
customer database;
Enterprise knowledge base;
Recommended workflow;
CRM system plug-in.

achieve effects

Systematically analyze customer consultation keywords, model preferences and historical behaviors, form user portraits, and provide sales personnel with recommended models, communication priorities and reference skills.

Automation of e-commerce order processing

business needs

Non-standardized orders often describe product specifications in natural language, and manual sorting is prone to omissions and entry errors.

use components

Workflow;
Big language model;
conditional judgment;
Data processing node;
Order system and supply chain plug-ins.

achieve effects

The system automatically reads order descriptions, extracts material, size, quantity and customization requirements, converts unstructured text into structured orders or BOM lists, and then transmits them to the supply chain system, thereby reducing manual entry and improving order processing efficiency.

Personalized reading assistant

business needs

Users hope to record reading progress, reading notes and interest preferences for a long time, and quickly restore historical information in subsequent conversations.

use components

Agents;
Database;
Long-term memory;
Prompt words;
Recommend workflows.

achieve effects

After the user enters the reading bibliography, page number and notes, the system saves the information to the database. The next time the user asks, the agent can resume reading progress and provide content review or extended reading suggestions based on historical notes.

English teacher

business needs

Learners need to conduct oral practice, grammar correction and situational dialogue at any time.

use components

Agents;
Big language model;
Role prompt words;
Voice tools;
Study record database.

achieve effects

Agents simulate English teachers, conduct English dialogues with users in different scenarios, and adjust the difficulty according to the user’s level. The system can also save common mistakes and learning progress and provide personalized practice suggestions.

Content Creation Assistant

business needs

Operations personnel need to quickly complete copywriting conception, content rewriting, title generation and multi-platform style conversion.

use components

AI applications;
Big language model;
Prompt word template;
Workflow;
knowledge base.

achieve effects

After the user inputs the theme, audience and content requirements, the system generates outlines, first drafts, titles and different platform versions according to the preset workflow, and maintains a unified expression style based on the brand knowledge base.

Coze platform advantages

Low development barriers

Coze provides a visual development environment. Users can complete application construction through form configuration, natural language prompts, and drag-and-drop nodes without having to write complete large model calls, tool scheduling, and process control code from the bottom.

This allows products, operations, teachers and business personnel to also participate in AI application design.

Flexible workflow orchestration

Workflow can split complex tasks into multiple nodes and support conditional judgment, looping, code processing, model calls, and plug-in calls.

Compared with a single model Q & A, workflow has the following advantages:

Clear business steps;
Input and output are clear;
The implementation process can be observed;
Complex logic can be controlled;
Processes can be reused.

Strong plug-in expansion capabilities

The big language model itself cannot directly obtain all real-time information, nor can it directly operate enterprise systems. Plug-ins enable agents to connect to search, weather, maps, email, image generation, and internal corporate APIs.

The official open source document divides plug-ins into official built-in plug-ins, custom plug-ins, and commercial plug-ins.

Support enterprise-specific knowledge base

The knowledge base allows agents to answer questions based on the enterprise’s own data and solve problems where common models do not understand internal business knowledge.

Its value is mainly reflected in:

Unified enterprise knowledge portal;
Reduce duplicate consultations;
Improve the accuracy of professional answers;
Shorten employees ‘time to find information;
Support questions and answers on products, systems and technical documents.

Supports databases and long-term memory

Coze can transform key information in conversations into structured data and reuse it in subsequent tasks.

This allows the agent to:

Remember user preferences;
Track task progress;
Preserving historical status;
Provide continuous services;
Achieve personalized recommendations.

Support fast debugging and iteration

The platform provides real-time preview and debugging areas. Developers can repeatedly test prompts, plug-ins, knowledge bases, and workflows before release, and modify configurations in a timely manner based on the results.

This development method shortens the cycle from creative verification to application launch.

Support multiple integration methods

Coze’s agents and applications can run independently or access existing websites, enterprise management systems and other software through APIs and SDKs.

This makes the platform suitable for quickly creating independent AI applications and adding AI capabilities to existing systems.

Support privatization deployment and secondary development

Coze Studio is an open source implementation of Coze’s core capabilities that can be deployed locally or on enterprise servers and allows development teams to customize based on source code.

This ability is suitable for:

Enterprises with high data security requirements;
Units that need to operate on the intranet;
Projects that require access to private systems;
R & D team that needs to modify platform functions.

It should be noted that privatization deployment still requires security configurations in terms of account permissions, interface security, data isolation and operating environment.

Coze platform usage process

Coze official website version (cloud version)

Create an agent

Log in to the button platform, find and click the Create button on the main interface to officially start the personalized construction process of the agent.

coze.cn

扣子 - AI Agent智能办公平台 - 扣子用AI重塑生产力与工作效率

扣子(Coze)是基于AI Agent的智能办公平台，提供AI写作、PPT生成、网页开发与AI设计等一体化生产力工具，扣子致力于为用户与企业提供高效、安全、便捷的智能办公新体验。

https://www.coze.cn/space

Give the agent a unique name, briefly describe its functional positioning and service scenarios, and clarify its role value.

Support the system to automatically generate avatars with various styles, and can also upload brand logos or exclusive images designed by individuals, allowing the agent to have a unique visual identity.

After confirming that the basic information is correct, click Confirm, and the system will jump to the core page of the agent choreography, and then you can carry out logical configuration and function development.

Write prompt words

The core of configuring the agent is to write a prompt word (Prompt), which defines its own human configuration and underlying reply logic. An accurate and clear prompt word can make the agent’s behavior pattern and output results more in line with our expectations, thereby efficiently completing specific tasks.

Add a knowledge base

The native capabilities of the basic model are not enough to support the complex needs of government investigation and business scenarios, and skills and knowledge bases need to be actively added to the agent.

This step can effectively break through the knowledge deadline and inherent ability boundaries of the model, allowing the agent to have advanced capabilities such as real-time information acquisition and tool invocation, and solving practical problems that cannot be completed solely by relying solely on large models.

Choose here to add a knowledge base

At this time, we do not have a knowledge base yet and need to create a new knowledge base

Here we choose to create the button knowledge base, select the text format, and fill in the relevant name and description

Select the import type of local document and change the icon, click Create and Import to create a new knowledge base

Then, we need to upload relevant files, wait for the analysis to be completed, and click Next

Here we select the default settings and next

You can see the segmented preview results here. Click Next after confirming that there is no error.

This will be submitted to the server for processing. No need to wait, just click Confirm

If necessary, you can add new content to this page. Here we choose to add it to the agent

add workflow

Select Add Workflow here

Because we don’t have a workflow yet, we need to create one, so here we choose to create a dialogue flow

You need to add the name and description of the conversation flow here

Here, the dialogue flow is built through visual operations

First create a knowledge base retrieval node, and use the mouse to connect the start node and the knowledge base retrieval node

Click the Knowledge Base Search node, select USER_INPUT of the start node from the input variables, and configure the existing Knowledge Base.

Add large model nodes and connect knowledge base retrieval nodes with knowledge base retrieval nodes and end nodes

Large model nodes require two input variables, USER_INPUT at the beginning and outputList for knowledge base retrieval

This node also needs to configure system prompts and user prompts

System prompt words and user prompt words require input variables referenced through {{}}, for example

用户问题：
{{input1}}

知识库检索结果：
{{input}}

请根据以上知识库检索结果回答用户问题。

Select the output variable of the end node, select the output of the large model node, and add the answer content {{output}}

Click Trial Run to test

Click on the upper right corner to publish the conversation stream and add it to the current agent

debugging agent

In the preview and debugging area on the right, talk to your robot like a real person. Through the actual interaction process, comprehensively test its answer accuracy, logical consistency and whether it meets your business expectations.

Release agent

After completing all testing and parameter configuration, just click the “Publish” button to officially deploy the agent. The system will automatically process back-end configuration to bring your agent online on the target channel and start providing services to users.

Here we choose to confirm

You need to select the publishing channel here. Button Store will be selected by default. You can do it by default.

After waiting for review, you can publish it successfully. After it is published successfully, you can search it in the work community

Coze Studio (open source version)

installation environment

The first thing you need to do is install Docker Desktop

docker.com

Docker Desktop: The #1 Containerization Tool for Developers | Docker

Docker Desktop is collaborative containerization software for developers. Get started and download Docker Desktop today on Mac, Windows, or Linux.

https://www.docker.com/products/docker-desktop/

Docker Desktop: The #1 Containerization Tool for Developers | Docker

Then download the source code for Coze Studio from Github

github.com

GitHub - coze-dev/coze-studio: An AI agent development platform with all-in-one visual tools, simplifying agent creation, debugging, and deployment like never before. Coze your way to AI Agent creation.

An AI agent development platform with all-in-one visual tools, simplifying agent creation, debugging, and deployment like never before. Coze your way to AI Agent creation. - coze-dev/coze-studio

https://github.com/coze-dev/coze-studio

GitHub - coze-dev/coze-studio: An AI agent development platform with all-in-one visual tools, simplifying agent creation, debugging, and deployment like never before. Coze your way to AI Agent creation.

After unpacking everything, open the terminal under coze-studio\docker and type

docker compose up -d

Used to start services

Register an account, visit http://localhost:8888/sign, enter your username and password, click the register button, and then you will come to this page

configuration model

Get vector model

First, download ollama to invoke the local model. The download link is

ollama.com

Download Ollama on Windows

Download Ollama for Windows

https://ollama.com/download/windows

Then open the terminal and enter

ollama pull bge-m3:567m-fp16

This command is used to download the local model and view it after downloading:

ollama list

Get the language model

First log in to the Alibaba Cloud Refining console

bailian.console.aliyun.com

大模型服务平台百炼控制台

百炼控制台是阿里云大模型服务平台，提供AI模型训练、部署、推理一站式服务，支持多种大模型框架，助力企业快速构建AI应用。

https://bailian.console.aliyun.com/

Then apply for an API Key

bailian.console.aliyun.com

大模型服务平台百炼控制台

百炼控制台是阿里云大模型服务平台，提供AI模型训练、部署、推理一站式服务，支持多种大模型框架，助力企业快速构建AI应用。

https://bailian.console.aliyun.com/cn-beijing?tab=model#/api-key

Copy key and request address

Visit Coze Studio 管理后台 and choose Qwen family down

Fill in the model display name and model type, paste the key and address, and click Save

Then fill in the following information, then swipe down and click Save Configuration

It prompts us to restart the server. We open the terminal under coze-studio\docker and enter

docker compose restart

At this time, the relevant services will be restarted

The subsequent configuration process is similar to the cloud version, but don’t forget to modify the model

Coze platform application scenarios

education and teaching

Suitable for building:

AI teaching assistant;
Speaking English sparring partner;
Course materials question and answer assistant;
Homework guidance assistant;
Learning planning tools;
Teaching content generation and application.

Agents, knowledge bases, prompt words and learning record databases are mainly used.

enterprise customer service

Suitable for building:

Pre-sales consulting robot;
After-sales service assistant;
Order Inquiry Assistant;
Product use instructions;
Manual customer service assistance system.

It mainly uses knowledge bases, business plug-ins, workflows and APIs.

enterprise knowledge management

Suitable for building:

Question and answer on employee system;
product documentation assistant;
Technical support knowledge base;
Project data retrieval;
Training assistant for new employees.

It mainly uses knowledge base, RAG, rights management and dialogue interface.

content creation

Suitable for building:

Copy generation;
Title generation;
Content rewriting;
Report writing;
script generation;
Brand content assistant.

It mainly uses large language models, prompt words, workflow and brand knowledge base.

office automation

Suitable for building:

Organizing meeting content;
Email generation and classification;
Form data processing;
Automatic report generation;
Task notification;
Daily assistant.

Mainly use workflows, plug-ins, code nodes and databases.

E-commerce and operations

Suitable for building:

Product description generation;
User comment analysis;
Smart shopping guide;
Order information extraction;
customer service automation;
Marketing campaign content generation.

It mainly uses agents, workflows, commodity knowledge bases and e-commerce system plug-ins.

marketing and sales

Suitable for building:

Customer portrait analysis;
Sales speech recommendation;
Market research;
Collate information on competing products;
clue classification;
Marketing content generation.

Mainly use databases, search plug-ins, workflows and large language models.

Personal efficiency tools

Suitable for building:

Personal assistant;
Reading record assistant;
Travel planning;
Schedule management;
Information summary;
Personalized learning assistant.

Mainly use agents, plug-ins, long-term memory and databases.

Scene selection reference

Requirements Characteristics	Recommended Coze Capabilities
needs to complete tasks via chat.	Agent
requires fixed inputs, outputs and interfaces	AI Applications
requires multiple determination steps to be performed	Workflow
requires real-time information	plug-ins and tools
needs to answer the	knowledge base and RAG
needs to remember user preferences and historical status	database and long-term memory
requires access to existing business systems	APIs, SDK, and custom plug-ins
requires local data storage or deep customization of	Coze Studio Privatization and deployment of
requires rapid online experience and completion of prototype	Coze official version of
requires research on open source implementation or modification of platform functions	Coze Studio
has no deployment operation and maintenance conditions	Coze official version
requires long-term running on the corporate intranet	Coze Studio

Suggestions for selecting official website version and open source version

Suitable for selecting the official website version of Coze

This is learning Coze for the first time and I hope to quickly familiarize myself with core functions;
Used for classroom demonstrations, course experiments and personal exercises;
A workable prototype needs to be made as soon as possible;
Failure to meet server deployment and system operation and maintenance conditions;
Hope to directly use the online capabilities and ecological resources provided by the official website;
The project does not involve highly sensitive internal data for the time being.

Suitable for choosing Coze Studio

Needs to be deployed in a laboratory, school or enterprise server;
Need to control the data storage location and operating environment;
Need access to internal databases, models or business systems;
Need to modify the platform source code;
Need to study the implementation methods of agents, workflows, RAG and plug-ins;
The team has Docker, server and software development capabilities.

summary

Coze’s core value is not just to provide dialogue with large language models, but to integrate models, prompt words, agents, workflows, plug-ins, knowledge bases, databases and publishing capabilities into a complete low-code AI application development platform.

From the perspective of product capability architecture:

The large language model provides understanding, reasoning and generation capabilities;
Agents are responsible for receiving user goals and organizing task execution;
Workflow is responsible for stably orchestrating complex business logic;
Plug-ins and tools are responsible for connecting external services;
The knowledge base provides professional and private knowledge for applications;
Databases and long-term memory are responsible for maintaining business status;
Visual workbench, debugging and publishing capabilities support the entire application development process.

In terms of usage forms, Coze official website version and Coze Studio respectively meet different needs:

Coze’s official version focuses on out-of-the-box use and is suitable for beginners, course teaching, personal experience and rapid prototype verification;
Coze Studio focuses on privatization and independent expansion, and is suitable for technical research, enterprise intranet deployment and deep secondary development.