You need to enable JavaScript to run this app.
DataAgent (OnPremise)

DataAgent (OnPremise)

  • Documentation
    • DataAgent (OnPremise)Intelligent Analysis Agent User GuideBest PracticesConfiguration and Tuning PracticesBest Practice: Intelligent Analysis Agent Configuration and TuningPrompt Configuration Practices
Copy page
Download PDF
Best Practice: Intelligent Analysis Agent Configuration and Tuning
Prompt Configuration Practices
Copy page
Download PDF
Prompt Configuration Practices

Functional principles

The prompt is a directive or requirement that applies globally to the intelligent agent. Every user's query will include the prompt, so it has a significant impact on the overall effectiveness of the intelligent agent. It is typically used to define the intelligent agent's role, general response logic, and related scenarios.

General configuration approach
  1. In most scenarios, prompt configuration is not required. Global prompt configuration for the intelligent agent may unexpectedly worsen the quality of many responses. Exercise extreme caution when adding global prompts.
  2. Do not mention or reference table names in the prompt, as the model does not know the dataset names. Mentioning table names only has a negative impact on optimization.
  3. There is no need to mention or use chart-related expressions in the prompt, as such expressions are ineffective.
  4. Default filter conditions should be implemented through filter configuration on the dataset from an engineering perspective. Do not make such requests in the prompt configuration.
  5. For data formats, do not require the model to implement them through prompt configuration. If the format is not supported by engineering or there are problematic cases, you can resolve this by submitting a feature request to the BytePlus product and research team.

Configuration examples

Clearly specify the logic for returned result content, data scope, data units, and more

Type

Prompt

Note

The current user is an administrator. Interpret and address issues from the perspective of a senior manager. When asking about contracts, orders, shipments, revenue confirmation, payments received, and similar issues, in addition to answering the main data metrics, also provide contextually relevant information such as customer details, sales department, and other information related to the main issue, as well as other metrics related to the main indicators, such as sales volume, power, amount, and more.

Note

The terms 'loss-making Sales representative' and 'loss-making contract' have the same meaning. Answer using the same logic, and include contract number, sales department, Sales representative, contract gross profit, and other information in your response.

Note

When answering questions related to "each*", group and aggregate based on the field name that follows "each".

Requirements

If the question ends with "how many people are there in different employment types", select the "employment mode - currently employed workforce utilization" field as the dimension for the query; if the question ends with "what is the proportion of non-frontline", select the "employment mode - non-frontline workforce proportion utilization" field as the dimension.

Requirements

For monetary-related metrics, convert the unit from yuan to 10,000 yuan, 100 million yuan, and so on according to the data size.

Requirements

The query results for all metrics must be accurate to two decimal places.

Indicate the business meaning

Type

Prompt

Note

TOPCon, PERC, and HJT refer to technology types, not primary financial categories.

Common issues

Prompt configuration did not take effect

Prompt configuration example

Issue phenomenon & possible causes

Optimization approach

A "Note" type prompt was configured in the intelligent agent: The name of the primary financial category is composed of "product specification + structural characteristics + technical route".

When using the intelligent agent for data queries, it was found that this rule did not take effect. When querying metric data for a certain primary financial category name, the data matching was incorrect.
Possible causes:

  • The business meaning of this prompt rule is relatively complex.
  • During recall before table selection in the intelligent agent, the prompt rule is recalled last. Before that, more suitable datasets have already been found based on recall results from datasets and knowledge, so this prompt rule did not take effect.

For prompts with relatively complex rules, it is recommended to configure semantic models for each dataset field involved in the business meaning within the intelligent agent's datasets.

Last updated: 2026.06.02 14:59:30