AI Agent
Last updated
Last updated
Currently, this feature is only available for web chat and WhatsApp. Each use consumes one AI credit.
This mode significantly reduces the workload of developing conversational Q&A flows. For our web interface, we offer two modes:
Suggestion Mode: The AI will not directly reply to visitors but will provide suggested search items in the chat.
Automatic Reply Mode: The AI will directly answer the visitor's questions.
Regardless of the mode, a notification will be sent immediately to online customer service agents for the first question. Additionally, a notification will be sent when the intent is not matched. Assigned agents are automatically distributed based on an automatic load balancing system, considering open chat windows.
Input:
Web Page: Supports text.
WhatsApp: Supports text, images, and voice messages (approximately one minute in length, file size should generally not exceed 5MB).
How to Use:
First, navigate to the domain to enable the functions you need.
Document Q&A:
This is the simplest method. You upload files, but note that in sitemap mode, only static web pages will be captured. For single dynamic web page capture, the recent version has the crawler user-agent preset to DMflowBot, and it will crawl according to the rules of robots.txt. If you are using Cloudflare and have WAF (Web Application Firewall) enabled, please disable it or allow DMflowBot to crawl. Please crawl your own web pages.
Form Q & A:
A is like using a database. You set the schema first. However, because the product itself is an adapter (converting SQL to the system's search method), it does not operate on a real database, so there are no issues like SQL injection.
After designing, please go to the dashboard to add settings. When the robot adds data, it will change according to your settings. If you do not add settings, it will automatically set according to the fields.
Chat Character:
Finally, go to the Chat Character settings. Please refer to Chat Character.
Start Setting Up an Agent:
Currently, agent settings are divided into precondition settings (filtering user information), intent settings, and response settings.
Filtering user information is not yet available and is mainly used to determine the user's chat rounds and whether the user's information meets the requirements (e.g., whether the email has been verified after being entered).
Intent settings are the main focus, and this tutorial will focus on this aspect.
Intents are primarily determined by descriptions. Visitor information is currently preloaded, and memory modules or database calls may be added in the future to make intent judgment more accurate.
"Allow Shortcuts" means you can use "/" followed by a name and a space, and then the conversation, to skip intent recognition.
For example: /query information I want to ask what your company's unified business number is.
Modes are divided into agent response and direct response.
Direct response means that when the intent is recognized, the message you have set will be replied directly.
Agent response means selecting your chat character, and when the intent is hit, it will switch to that agent for response.
"Default Open Tools" means that tools must be called every time the agent is called.
"Default Open Thinking" means that the agent will think before calling tools.
A point to note here is "thinking".
When you enable "thinking", it is recommended that your prompts should not be role-playing like "What kind of role are you?" Instead, write what you want to do directly in the prompt. For example: "Before adding user data, you should first check if the user exists." This kind of command writing prompt is sufficient.
Testing
Each AI conversation will provide the full content of "thinking", automatic expansion of "knowledge" queries, and "followup_questions" to recommend the user's next conversation steps. Other features include the ability to display insertion messages for form insertions, and form searches will display the form content that the visitor can query (currently, expressions like "not equal to" are not supported). Clicking "More" will give you a direct view.
Next to the visitor, there is a magic wand that can translate the language according to the language selected by your browser. The conversation menu from left to right includes emoji, note (add shared notes, use @ to call other specialists), and magic (currently only provides translation functions, and if the other party's locale has a value, it will default to the other party's language).
The free version of AI translation provides 10,000 characters per person per month, and the light version provides 100,000 characters per person per month. Character count is limited by the number of input characters. This is sufficient for general customer service.
Regarding permissions, insertion by an agent is allowed (if you insert from the dashboard without permission, it will fail), but query permissions must be set correctly. Please refer to for details.
After clicking "Add Settings," you can set default values, required fields, etc. Note that Upload currently only supports UI functions and does not support any chatbots. If you want to trigger a workflow after adding, please set it in Button. For detailed triggers, please refer to or .
Go to Settings -> Embed Webpage settings. Pay attention to the following: enable "Share Open", enable "Enable Embed Webpage Open", and add "" to "Domain Restriction". After adding, open the share link below to test.