Using Coze, I created a simple WeChat assistant that connects to a public account for conversation. Currently, it can answer questions based on stored articles, and you are welcome to try it out.
Effect Testing
Question List:1. What is the main purpose of this public account?2. Are there any recommended articles?3. Are there any articles related to AI?4. (AI Article) What is it mainly about?
Implementation Method
1. Knowledge Base
Add public account articles to the Coze knowledge base. To facilitate searching for specific article titles, the knowledge base is recorded in a structured Excel format.
2. Workflow
Design the conversation flow based on user input for intent recognition. If the question involves public account content, search for related content in the knowledge base (RAG) and provide the retrieval results to the large model for summarization and response; if not, respond using the large model as a fallback.
Below is the entire workflow arrangement.
3. Agent Definition
# Role
You are the intelligent chat assistant for the public account "Continuous Output of Shuoen", answering questions for followers of the public account. Shuoen focuses on sharing trends and insights related to AI, Web3, and finance, as well as reflections on life and work, aiming to organize its knowledge thinking system while enhancing readers' understanding of technology and finance.
Your task is to accurately understand the user's question intent, summarize the content related to the user and the question from the given article, and respond.
## Skills
### Skill 1: Filter Related Articles
1. Carefully analyze the keywords in the user's input.
2. Traverse the article list to find articles with keywords in their titles or content.
3. Return a list of article titles in a format with links.
### Skill 2: Summarize Article Themes and Views
1. Carefully analyze the user's intent and the keywords of interest.
2. Find answers to the user's questions from the articles or summarize content related to the user's keywords.
### Skill 3: Guide Questions
1. Accurately analyze the user's question intent; if the question is unrelated to the public account's theme, cleverly guide the reader to ask questions related to the author's articles and themes.
2. When the user's question is vague, actively ask for more details to better guide the reader to ask questions about the article content.
## Limitations:
- Only answer questions related to the content and themes of the public account "Continuous Output of Shuoen", and refuse to answer unrelated topics.
- All output must be organized according to the given format and cannot deviate from the framework requirements.After completing the above process, use Coze’s built-in publishing function to connect to the public account backend to enable dialogue between the public account and the agent.
Upon following the public account, you can receive the following automatic replies.
As an attempt to use intelligent agents, currently only a small number of articles are in the database, and the definition and workflow design of the agent still have significant room for optimization.
Through the creation of this agent, the biggest realization is that the agent can cleverly integrate the powerful capabilities of large language models with traditional solutions to real-world problems, demonstrating results beyond expectations. This reminds me of the launch of the first generation iPhone. At that time, the iPhone did not adopt entirely new technology but integrated mature technologies such as capacitive touch screens, communication technologies, and graphical interfaces, creating a revolutionary product. Similarly, the agent achieves impressive results by combining large models with traditional programming methods.
Of course, this is just a starting point. In practical applications, how to more accurately split user intent and design a more efficient and cost-effective workflow requires extensive trial and error and optimization.
After all, I do not want to receive emails at midnight saying that token consumption exceeds the limit.
Only through continuous exploration and improvement can we fully unleash the potential of intelligent agents and promote their widespread application in more fields.
– THE END –