Natural language instructions in Spotter

The natural language instructions feature allows you to provide the Spotter system with global rules that it should follow across all relevant queries. Unlike Spotter context, which is tied to a specific reference question, a natural language instruction is a standalone directive designed to teach the system a core concept, a default behavior, or a specific data nuance. These instructions are sent along with the user’s query to guide the system toward a more accurate and consistent response.

These instructions are globally applicable and influence all queries made by any user on the data model, unlike other coaching methods that also have a user-level scope.

Natural language instructions are in beta and off by default. To enable this feature, contact ThoughtSpot Support.

Access and permissions

To add or manage natural language instructions, you must have data model editing access.

Key use cases

Natural language instructions are a versatile tool for shaping how the system interprets queries and presents answers. The primary use cases are:

Data filtering and refinement

Apply automatic filters to ensure analyses are consistent and accurate. This can be broken down into two main types:

Applying conditional filters

Teach Spotter to apply different logic based on the user’s query.

Example

To use different fulfillment centers based on the product line, you can instruct: "For queries about 'Electronics' sales, filter for Fulfillment_Center = 'FC-West'. For 'Apparel' sales, filter for Fulfillment_Center = 'FC-East'".

Applying default filters

Set up a rule that always applies a filter to clean up data.

Example

To ensure internal test data is always excluded from revenue analysis, you can instruct: "When calculating revenue, always exclude transactions where Account_Type = 'Internal Test'".

Column selection and prioritization

Guide the system to use the correct columns for analysis, filtering, or aggregations.

Example

If a query for "market" could ambiguously use a City or Canvas Market column, you can direct the system with: "Prefer Canvas Market column over City column for filtering market values".

Example

To ensure counts are always accurate, you can specify the column to use for counting: "For counting customers, always use the unique count of ‘Customer Cred ID’".

Answer enrichment

Automatically add more detail to answers to make them more comprehensive without requiring the user to ask for every column explicitly.

Example

If a user asks for basic provider information, you can enrich the answer with more detail using the instruction: "Provider demographic info is name, address, tax info, class and state code".

Date handling

Manage how the system handles date and time analysis, especially when a user omits a time frame.

Example

To prevent expensive, wide-ranging queries and align with common user preferences, you can set a default time period: "If a query does not specify a time period, apply a default filter for the 'last 30 days'".

Give instructions in Spotter

To give instructions in Spotter, follow these steps:

  1. Open Spotter and ask a question. For example, you could say, “Show me the sales for the last month.”

  2. Let’s say your company defines “last month” as the last 30 days. If Spotter answers your question with the date filter “last month,” you can add context to correct its answer.

  3. Click the information pane icon in the top right corner of the conversation. The instructions pop-up appears.

  4. Click the + icon to add an instruction. In the text box that appears, type your instruction. For example, you could say, “When I ask for last month, use ‘last 30 days’ as a filter.”

  5. Click the check mark to save your instructions, and select Close.

  6. When the pop-up closes, Spotter displays an acknowledgement that Instructions have been updated. Click Regenerate last answer to see the updated answer to your question and verify that Spotter has understood your instructions.

Best practices for writing instructions

Be specific and direct

Write your instruction as a clear command to ensure the system applies your rule correctly and predictably. For example, "Exclude all sales from 'Internal Test' accounts" is better than "I don’t usually want to see our internal test accounts in reports."

Pro tip

In many cases, using preferential language like "Prefer A over B" can be more effective than a hard command like "Use A".

Group instructions by concept

Instead of creating many separate instructions for small, related rules, it is better to group them. Focus on teaching all the logic related to a particular business concept in a single, well-structured instruction. This helps the system understand the full context and know which rules to apply when a user’s query mentions that concept.

Match your data’s language

Write instructions using the exact names for columns and the precise values stored in your data. This removes guesswork for the system and leads to more predictable results.

Example

Instead of a generic (less effective) instruction: "Exclude canceled orders when calculating total revenue", use a more precise (more effective) instruction: "Exclude orders where order_status = 'CANCELLED-USER' when calculating total revenue".

Limitations and key considerations

No automatic override behavior

A natural language instruction does not automatically override a user’s direct query if they conflict. Instead, both the user’s query and the instruction are sent to the Model, which can lead to unpredictable results or confusing answers. This makes writing clear, specific instructions critical.

Not recommended for mathematical formulas

Natural language instructions are not well-suited for teaching the system mathematical formulas or complex calculations. For these scenarios, other coaching methods, such as providing a reference question with natural language context, are strongly recommended.

Avoid conflicting coaching

The instructions you provide must be coherent with each other and with other coaching assets like reference questions and business terms. Contradictory information will confuse the model and lead to unpredictable results.

Unsupported capabilities

To set clear expectations, please note that several advanced “agentic” behaviors are not currently supported by natural language instructions. These include:

  • Enforcing specific chart types or visualizations.

  • Prompting the user with follow-up questions.

  • Defining conditional query logic (for example, “if the first query returns no data, run a second query”).

  • Question deflection (preventing the system from answering certain questions).

Was this page helpful?
×