Administrators can create multiple agents based on business requirements to meet the daily usage requirements of members.
The following figure shows the effect after agent creation.
Before creating an agent, you must have completed the following preparatory work to ensure smooth configurations:
Model preparation: Choose Admin Console > Model, and add and test available large language models (LLMs).
Data preparation: Choose Admin Console > Data, and add the data/dashboards required for agent analysis.
Knowledge preparation: Choose Admin Console > Knowledge, and add the knowledge library required for agent analysis.
Business combing: Clarify the agent's business scenarios, capability boundaries, and user requirements to ensure that the configurations align with actual businesses.
Choose Admin Console > Agent, and click the Create Agent button in the upper right corner to enter the creation page.
The following two creation methods are supported:
Encapsulated Agent: Select your own model and skills to build an agent within the platform.
Integrated Agent (Coming Soon): Connect to an agent already published on an external platform.
After entering the agent configuration page, click the edit button next to the agent name to modify the following basic information:
Agent Name: Customize the agent name to clearly reflect the business scenarios.
Description: Briefly describe the agent's purpose and capability boundaries.
Icon: Select a preset icon for easy user identification.
Label: Select one or more suitable label categories for the agent.
Avatar: Based on the administrator's settings in Common Agent Management, an AI-generated random avatar or a custom avatar will be used by default. Developers can also regenerate one.
If you need to exit the current configuration page before publishing the agent, save the configured content promptly to avoid loss.
Before publishing an agent, developers can set its visible range. The following three visible ranges are supported:
Visible Only to Creator
Visible to Some Users
Visible to All Users
(2) If an agent's visibility modified by a super administrator has exceeded the permission scope of the agent's current editor, the current editor can unpublish the agent, adjust the visibility, and then republish the agent.
(3) The current editors of the agent are its first-order administrators.
After the agent is configured and tested successfully, click the Save and Publish button in the upper right corner. You can adjust the visible range again.
You can select from the following publishing methods and confirm the publishing:
Publish the agent to the user workbench for business personnel to use.
Generate an independent URL for the agent to be quickly copied and shared externally.
You can publish an agent to one or more resources under FineBI > Directory. When users access the corresponding resources in the directory, they can quickly initiate an agent chat through the sidebar on the page.
(1) Unpublished agents are in the draft status and can be viewed and edited only by administrators.
(2) A single agent can be published to multiple resources in the directory on the FineBI platform, but a single resource in the directory can be associated with only one agent for publishing
To unpublish a published agent, go to the agent configuration page and click the Unpublish button. After unpublishing, common members will not be able to find this agent via search on the homepage, and any independent URL shared before will become invalid.
(1) Any configuration changes for a published agent need to be republished before taking effect.
(2) A published agent cannot be deleted. Unpublish it first before the deletion.
(3) You are advised to create agents by business line/scenario to prevent a single agent from having too complex capabilities. Otherwise, the response efficiency will be affected.
(4) As the core of agent behavior standardization, the system prompt needs to be clear without ambiguous expressions.
(5) The skill configuration needs to match the business scenario. Avoid adding unrelated skills that may interfere with the agent's intent recognition.