Coaching Spotter

Spotter coaching provides you options to map data to your business questions and terms.

Why coach Spotter?

Spotter is intelligent, but it doesn’t automatically understand the unique terms and rules of your specific business. Large Language Models (LLMs) like the ones Spotter uses are coached on vast amounts of public data and general language.

This means they don’t automatically know things like:

What "active customers" specifically means for your business.
- For instance, if you ask for "active customers," a general AI might assume this means "customers with a recent purchase." But in your business, "active customers" might specifically mean "customers with a current subscription, excluding those on internal trial accounts."
Whether "last month" in a particular context refers to an end_date or a closed_date in your dataset.
How to understand complex requests like "premium accounts in North America excluding internal partners."

That’s where coaching comes in. It provides Spotter with the necessary information about the semantics and context of your organization’s data.

Table 1. Benefits of coaching Spotter
Benefit	Explanation
Higher accuracy	Spotter learns to define terms like “booked revenue” or “churned user” based on your business logic, not general assumptions.
Less dependency on analysts	Business users won’t need to memorize specific data field names or complex filter logic to get answers.
️ Consistent definitions	Key metrics and filters are applied the same way for everyone, ensuring consistency across teams.
Faster decisions	Spotter answers more questions correctly the first time, reducing back-and-forth and speeding up insights.
Lower training burden	Users can ask questions naturally, without needing to learn new software or complex data structures.

Grant coaching access

You can now delegate Spotter coaching responsibilities without granting data model editing permissions. This feature is available for all data model editors and administrators. Granting coaching access allows your power users to refine coaching on a data model.

For more information on how to grant coaching access, see Grant coaching access.

Understanding coaching tools

Coaching is a multi-layered process. You must first create a strong metadata foundation before using the active coaching tools.

Metadata optimization

This is the most critical first step. This involves preparing your underlying data so Spotter can understand it better. Key aspects include defining proper column names, adding clear column AI context, and including relevant column synonyms based on your business use case.

This is required in all scenarios and should be addressed first. Well-defined metadata helps Spotter accurately identify and use the correct data fields when responding to user questions. For instance, if a column is named txn_dt but business users commonly refer to "transaction date" or "order date," renaming the column for clarity or adding these as synonyms (in case you don’t want to modify this) in your data model is a key metadata enrichment.

Similarly, for columns with indicator codes (for example, 1/0 or true/false), such as "valid_indicator_cd", it’s important to add clear AI context for the column —for example, "true means a valid transaction, false means an invalid transaction." This allows Spotter to interpret these codes accurately in business context, leading to more precise answers for your users.

For date fields, where a data model might have multiple date columns (for example, "order_date," "ship_date" "close_date"), providing clear AI context on each date column-- specifying which measures or metrics should be used with which date-- can help Spotter choose the right date column for each type of analysis, improving the accuracy of responses.

For more information on preparing your data for Spotter, see Spotter Model readiness.

Coaching tools

For coaching, the best practice is to start with the most foundational tool, Natural language instructions, to set global rules, and then use the other tools for more specific coaching.

Natural language instructions

These provide global rules to guide Spotter’s interpretation of a user’s query and the data model itself. Unlike other coaching methods that have a limited scope, these instructions are used while processing every relevant query from every user.

Best for: You can consider this as the most important do’s and don’ts for your new AI analyst. Set broad, consistent rules like applying default filters (for example, always excluding test accounts) to resolve ambiguity.

Reference questions

After setting your global rules, use these to teach Spotter "If a user asks X, you should answer with Y". This is enhanced with natural language context, which lets you explain why the answer is correct.

Best for: Frequently asked questions by your users can be added here so that the most common questions are answered efficiently. Additionally, you can coach complex, multi-step formulas or resolve ambiguity for specific common questions that a global rule can’t cover.

Business terms

This is your final, most specific tool. It should be used as a "last resort" to create a specific, reusable TML (ThoughtSpot Modeling Language) mapping for a term.

Best for: Use this feature to create simple, universally true definitions, such as mapping a value synonym (for example, "N.Am." → country = 'North America') or a very simple, universal formula.

Natural language instructions

Natural language instructions are directions you give to Spotter, either within a conversation or on a Model. These rules apply to all searches held on that data source.

How natural language instructions help Spotter

Natural language instructions are your strongest tool to coach Spotter-- unlike reference questions and business terms, they are referenced in every relevant search on the underlying Model. You can set rules to teach Spotter logic based on the user’s query, to set up rules that apply filters to clean up data, and to guide the system to use the correct columns for analysis, filtering, or aggregations.

Reference questions

Reference questions are a set of sample questions and their corresponding answers (in ThoughtSpot Search keyword language).

How reference questions help Spotter

The reference questions provided to Spotter help it in the following ways :

Spotter provides the verified answer when users ask the same question

The reference questions are designed to provide the exact same answer when users ask the same question. We recommend adding the most commonly-asked questions to reference questions so that business users get a verified answer that is curated by an analyst.
For the questions specified under reference questions, Spotter also remembers the visualization settings set by the analyst.

Reference questions are not designed to coach on visualization settings, as the specific visualization settings you apply (like chart type and axis setup) are only applied if the user’s question exactly matches the reference question. Even a one-word difference means those custom chart settings won’t carry over.

Spotter learns how different columns should be used to answer questions

An indirect benefit of providing reference questions is that it helps in warming up ThoughtSpot’s usage-based ranking systems. The reference questions help our system learn which columns must be selected in answering questions.
For newly-created data sets where the usage-based ranking system isn’t warmed up, Spotter might struggle to pick the correct columns if there are similar-sounding column names. Adding reference questions coaches Spotter to understand which columns should be used for answering questions.

Key considerations

New data sets: For new data sets, especially those without much usage history or existing Liveboards, reference questions are vital. They help "warm up" Spotter’s understanding of which columns are most relevant for analysis, especially if column names are ambiguous.
Avoid over-coaching: Don’t create too many reference questions that are extremely similar but with tiny variations in wording. This can sometimes confuse Spotter. Test Spotter with variations of a concept. If it understands them after coaching one good reference question, you don’t need to add redundant examples.

Coaching a sales metric

Let’s say you want to train Spotter on a specific question, “What are the sales for Eastern stores this month?” You want Spotter to understand this as Sum Order_Amount where region = ‘East’ and order_date = this month.

From this, Spotter might learn to associate:

“sales” with the Order_Amount column (and the general intention of total sales while referring to sales).
“Eastern stores” with the region= ‘East’ column .
“order_date” for queries referring to sales topics.

Once coached with this reference question, Spotter should ideally also respond correctly to variations like:

"Sales for the West this month?"
"How much did we sell in the North this year?"

Key point to note: Always test these variations. For terms that always need a specific mapping (like "sales" always meaning sum(Order_Amount)), using a business term is often a better approach.

Guidelines for reference questions

Add commonly-asked questions in reference questions

Spotter provides the answer you have curated when a user asks the same question. Hence, we recommend that you add the most commonly-asked questions by the business user as reference questions so that users get a verified answer every time they ask these questions.

Coached question templates

You should think of reference questions as coached question templates. For example, once you have provided “What are the sales for east this month?” as a reference question, then Spotter has the ability to generalize the learnings from this question to answer similar questions like “What are the sales for west this month?”, “What are the sales for north this year?”, “What is the quantity sold for east this week?”, etc..

Please note that the visualization settings provided in query coaching may not extend to these questions. Visualization settings don’t apply if the question is different from the reference question.

Ensure good column coverage

As mentioned above, Spotter learns how to use columns for analysis when they are used in reference questions. Hence, we recommend that you provide enough reference questions so that all the important columns in your Models are represented in the reference questions.
For newly created data sets with no Liveboards, the reference questions play an extremely critical role in coaching Spotter.

For details on how to create reference questions, see Creating reference questions.

Business terms

Business terms give you an option to create a mapping between your data and the business term/vocabulary used in day-to-day operations.

How business terms help Spotter

Business terms are a mapping of the data to the vocabulary used by the business users in day-to-day life. Every business has its own definition of various metrics and a very specific way for calculating these metrics. Business terms allow an analyst to add these mappings to Spotter to ensure that the LLM uses your definition of a metric instead of making assumptions to calculate these metrics.

We recommend using business terms for providing the following types of knowledge to Spotter:

Define how to calculate certain metrics: In order to answer questions, you sometimes need to generate calculated fields or metrics. You can use the business term to teach Spotter how to calculate specific metrics for your business. Spotter learns how you calculate specific metrics, and it has the ability to extrapolate these learnings to different scenarios.
Define synonyms for column values: Often, the same value might be referred to with different names. We recommend the use of business terms to define the synonyms. While Spotter has the ability to apply semantic matches based on publicly-known information, there are always some cases where you want to define synonyms or acronyms which are only applicable to your own business and cannot be guessed by a large language model.
Filters definition: Business terms can be used to define filters that should be applied for specific cases.

Guidelines for business terms

Our system suggests business terms to you when you are coaching Spotter for reference questions or correcting the answer during conversation. You must only add business terms which meet the following criteria :

Meaningful addition to Spotter knowledge: The first thing you should evaluate is whether adding the business terms will lead to a meaningful addition to Spotter knowledge about your data. Anything which is obvious or common knowledge can already be guessed by the LLMs as they are coached on a large body of public data sets. Hence, try to avoid adding the definition for commonly-available terms in business terms, as the LLM will already provide correct responses.

Business terms are most helpful for addition of definitions which are specific to your organization and cannot be guessed by business users. These are the business terms which will provide a meaningful addition to Spotter knowledge.

Note there are terms like “sales contribution” which may seem fairly obvious at first, however, when you deep-dive into specifics of calculating it, you will realize that the way your business computes this is very specific. Hence, we recommend that you provide some sample to Spotter to ensure these definitions are calculated according to your business requirements only.
Hold the same meaning in all contexts: Business terms, once defined, are considered to apply uniformly across all the question contexts. You must only use business terms for adding analytical definitions for the terms which hold the same meaning in all the different contexts it can be used for that Model.
You should avoid addition of business terms just to coach Spotter on handling date intents better because in most data sets there may be more than one date column, and once you have used business terms to define date intent (for example, this year → created_date.this year), then Spotter might start using the same definition when this year is used in a different context.
Business terms and their analytical definition are correct: You should only accept the business terms whose analytical definition (represented by ThoughtSpot keyword-based search tokens) is correct.

For details on how to define business terms, see Defining business terms.

For more information on your coaching strategy, see Coaching best practices.

Was this page helpful?Give us feedback!