AI-Native Database Service:Forecast Agent

Adjust the suggested ontology during creation

• All entity types and edge types are selected by default. A highlighted tag indicates that the type is retained. Click a tag to clear the selection, and click it again to select the type.

• When the agent group is created, only the currently selected entity types and edge types are used.

• Candidate entities are only a preview of the recognition results. You do not need to select them one by one.

• If a required type is missing, enter a name in Custom Vertex Type or Custom Edge Type and click Add. New types are selected automatically.

• If you return to the left pane and modify the source text or the creation goal, click Generate Ontology again and check the selection results.

Use short, unambiguous names and a consistent naming convention for custom types. For example, use Consumer and KOL for vertex types, and CARES_ABOUT and INFLUENCES for edge types. Add a type only when a key business type is missing, and avoid creating many types with similar meanings.

Maintain an ontology after creation

1. Go to the agent group details page and click Ontology Management at the top of the page.

2. Switch between the Entities and Relationships tabs.

3. Click Add Entity Type or Add Relationship Type, enter a name, and confirm the operation.

4. Click an existing type card to add a description for the type. For an entity type, you can also specify whether the type represents a person or an organization.

5. To update the graph and members based on the latest ontology, click Rebuild Agents and wait for the task to complete.

Add a custom agent

If the generated group lacks a role that must be present, add an agent manually.

1. Click Agent List and then click Add.

2. Enter a name and describe the persona in Markdown.

3. Click Add in the form. After the agent is created, the member is labeled Manual.

Delete an agent

The system reduces noise and resolves ambiguity when it builds an agent group, but some agents may still not meet your expectations. In this case, delete the agents manually.

Click Agent List, move the pointer over the target agent, and then click the delete icon that appears on the right. After an agent is deleted, the system automatically removes the corresponding entities and relationships from the graph.

Edit a persona and start a temporary conversation

The system generates a persona for each agent based on the uploaded background materials. You can also modify a persona.

1. Click Agent List and select the agent that you want to edit.

2. On the persona page, click Edit and modify the persona in the text box. Use Markdown to obtain better results.

To verify whether a persona meets your expectations, click Chat in the upper-right corner of the agent persona page and start a temporary conversation with the agent.

Information diffusion

Enter the complete original text of the first release in the Initial Information field. Do not summarize conclusions for the group in advance. Specify the time, scope, target, and exceptions as clearly as possible.

Sample input

Initial Information: Starting September 1, 2026, the downtown area of the city runs a 30-day trial of odd-even license plate restrictions for fuel-powered passenger cars from 07:00 to 09:00 and from 17:00 to 19:00 on weekdays. New energy vehicles, ambulances on duty, and school buses are exempt. Residents can check whether a vehicle is restricted on a given day in the City Transport mini program. Violations receive only a reminder in the first week of the trial, and penalties apply from September 8.

Stance evolution

Both Stance Space and Opening Post are optional. Specify them when you already know the main stances, which makes later comparison easier. Leave them empty when you are not sure, and the system infers the stances from the topic.

Sample input

• Discussion Topic: Starting January 2027, should the company make every Friday a company-wide remote work day?

• Stance Space (optional): Support / Conditional support / Oppose / Undecided.

• Opening Post (optional): The company is evaluating a fixed remote work day. Discuss whether the policy should be adopted and which prerequisites are required, based on cross-team collaboration efficiency, employee commuting costs, customer response speed, and management fairness.

Multi-option consensus

Specify at least two candidate options. Describe each option along the same dimensions, such as target users, benefits, costs, and risks, instead of highlighting the advantages of only one option.

Sample input

• Decision Question: Which packaging version should the new product finally use?

• Option A (minimalist professional): White and platinum colors that emphasize a professional look; suitable for business users; production costs are roughly the same as the current version; the risk is weak visual recognition on social platforms.

• Option B (young and energetic): Highly saturated contrasting colors that suit younger users; high recognition on social platforms; production costs increase by 8%; the risk is that the established professional brand image may weaken.

• Option C (eco-friendly material): Recyclable kraft paper that emphasizes sustainability; production costs increase by 5%; the risk is that moisture resistance and a premium feel still require verification.

Fact correction

Enter the original claim that requires correction, and do not mix the clarification conclusion into the same field. When you compare different early feedback conditions, change only this setting in each run.

Sample input

• Disputed Claim: Online posts claim that the company will cancel all membership discounts in physical stores next month, that existing members will be able to use points only online, and that in-store purchases will no longer receive any discount.

• Early Feedback Condition: Natural control.

To compare how early public opinion affects the correction result, keep the disputed claim and the participants unchanged, and then run the deduction once with Preset Positive Buzz and once with Preset Negative Buzz.

Buzz and word-of-mouth influence

Keep the launch content as close as possible to the real announcement, and retain the price, benefits, and restrictions. When you compare different feedback conditions, do not change the content or the participants at the same time.

Sample input

• Launch Content: The brand officially releases the P1 portable projector at CNY 1,999. The product targets camping and rental scenarios and supports 1080p resolution and autofocus. Early adopters can return the product within 7 days without providing a reason. However, outdoor use requires a separate power supply, and some streaming platforms require a separate membership.

• Early Feedback Condition: Preset positive buzz.

For a controlled deduction, keep the launch content and the participants unchanged, select Natural Control, Preset Positive Buzz, and Preset Negative Buzz in separate runs, and then compare the three results.

Key figure intervention

You can specify one or more key figures. Describe the identity, source of credibility, main viewpoints, and suggested actions of each figure. A name alone reduces the explainability of the results.

Sample input

• Opening Post before the intervention: The platform announces that the merchant commission rate will increase from 20% to 25% starting next month. The platform states that the additional fees will fund traffic subsidies and after-sales guarantees, but some merchants are concerned that operating costs will continue to rise.

• Key figure 1: Li Chen, an industry researcher who has tracked local lifestyle platforms for years and contributed to several industry cost studies. His view is that the available data is insufficient to conclude that the commission increase will cause a large-scale merchant exodus. He suggests that the platform first disclose how the commission is used and provide merchant revenue estimates, and then observe the effect for one settlement cycle.

• Key figure 2: Wang Min, a merchant association representative who participates in platform rule discussions on behalf of small and medium-sized merchants. Her view is that the commission increase significantly reduces the profit of low-margin merchants. She suggests that the platform postpone the change and invite merchants of different sizes to assess the impact together.