Training for Better Answers

Database schema is crucial for understanding the structure of your database, but it might not fully convey the meaning and business logic behind it. Therefore, AI sometimes requires additional "Background Knowledge" to accurately answer your questions.

Consider running a social network. When asking:

What's the total number of active users last week?

It can be challenging for AI to define "active users" without specific context. This term might refer to users who logged in, made a purchase, or visited the website last week.

Think of AI as a new intern: incredibly intelligent but needing guidance to grasp the full picture.

Another example involves managing an ad platform similar to Google Ads, where advertisers pay per click. Your "Click" table, with columns "user_id", "ad_id", "cost", "click_time", records revenue per click in the "cost" column.

Asking:

What's the average revenue per user last month?

might stump the AI without knowing that "cost" represents revenue from clicks on an ad platform.

AI requires sufficient "Background Knowledge" and clear examples to provide accurate answers when the schema alone is insufficient.

If answers are unsatisfactory, inspect the business definition and relevant schema before adding examples. Use a model supported by your current product configuration and evaluate it on your own questions. Documentation and SQL examples provide context; they do not guarantee correct answers or permanent model learning.

Documentation and SQL Examples

The training data is divided into two parts: documentation and SQL examples.

To add new training data, click the Train button in the top right corner (Desktop App):

Documentation

Documentation consists of a list of documents that provide background knowledge of the database.

Consider an Ad platform database, for example, you might provide the following documentation:

• This database relates to an Ad platform. The "Click" table records all ad clicks, with the "cost" column reflecting the revenue generated by each click.
• Users own multiple applications, with the Click table linking to the Application table via the "application_id" column.
• Consequently, each user's total revenue is the sum of revenues generated from clicks across their applications.

Keep each definition concise and follow the limits shown in your current product version. Desktop and Website Chatbot controls can differ; the screenshots below illustrate their documentation workflows.

For Desktop App:

For Chatbot:

SQL Examples

SQL Examples comprise a list of SQL queries for the AI to learn from, effectively pairing "Questions" with "Answers".

For instance, to calculate a user's total revenue, you might add:

For Desktop App:

For Chatbot:

This approach helps the AI to accurately respond to similar queries, such as "Who had the highest revenue yesterday?".

Think of SQL examples as "Quick actions" for frequently asked questions. Instead of specifying tables and columns each time, adding examples allows you to simply ask the question.

For example, after adding a revenue-related example, there's no need to manually define tables and columns for similar future inquiries.

Where is the Training Data Stored?

Storage and processing depend on the product and deployment. Desktop uses local application storage; hosted Chatbot documentation and SQL examples are saved with its configuration. Relevant context can be included in AI requests. Local storage does not mean offline model processing. Review the security and data-flow documentation before adding sensitive information.

Check the controls available in your installed version before planning an import, export or shared-training workflow.

For non-technical users, consider requesting assistance from your technical team to add training data.

Prepare business context before training

Use the AI database query guide and editable context brief to document entity counts, relationships, valid codes, date boundaries and exclusions. Its active-user example shows how a correct SQL query differs from a plausible but wrong count. The worksheet is article material, not an automatic product import.

Validate training with separate test questions

Use the text-to-SQL acceptance test kit to separate development examples from evaluation questions. It includes synthetic data, eight known-answer cases and a blank scorecard. Adapt it to your supported database, verify the reference answers there, and keep evaluation answers out of the training context.