Semantic layer

The semantic model is not built by an AI guessing at your business logic. It is defined and verified by humans, capturing tacit knowledge — the kind that lives in spreadsheets, email threads, and the heads of your most senior analysts — and encoding it formally.

ThoughtSpot’s token-based approach makes this practical. Verifying a definition doesn’t require writing or reviewing SQL. Business users and data stewards can crowdsource definitions in plain language, without needing to understand the underlying database structure. This turns governance from an onerous upfront project into a living, evolving practice.

Most metadata systems were built for BI tools, not AI. Spotter semantics is different: the semantic model is stored in a format that AI agents can directly leverage, giving them the full context they need to resolve user intent correctly.

When Spotter receives a natural language question, it doesn’t just pattern-match against column names. It consults a structured representation of your business — metric definitions, join logic, fiscal calendars, row-level security rules, AI context — and uses that context to generate an answer that is grounded, not guessed.

This is ThoughtSpot’s most technically distinctive capability. Rather than letting an LLM hallucinate SQL, ThoughtSpot uses proprietary, database-specific query generation to ensure every answer is:

The result: agents can interpret intent probabilistically and return deterministic results. The question is flexible; the answer is rigorous.

Scenario: Maria is a regional VP of Sales preparing for a board meeting. She opens Spotter and asks: "What’s our Q1 pipeline coverage by segment compared to last year?"

Without a semantic layer, the AI would pick whichever column looks most like "pipeline" and apply whatever date logic seems reasonable. Maria might receive an answer — but she wouldn’t know if it matched the definition her CFO uses.

With Spotter semantics, the query is automatically grounded in Finance-approved pipeline definitions, the company’s fiscal Q1 dates, and Maria’s regional data access permissions. She gets a reliable, board-ready answer in seconds — no analyst in the loop required, no second-guessing the numbers.

Scenario: James is a senior data analyst at a retail company. His team is constantly fielding ad-hoc requests: "What’s the return rate for our new product line?" or "Which stores had the highest basket size last month?" Each request takes hours to translate into a correct, governed query.

With ThoughtSpot’s semantic model, James can encode the official definitions once — return rate = orders with a return flag / total orders, excluding cancelled orders — and expose them as governed, searchable metrics. You can define this context in your queries, as well as adding it to your Models before searching. Non-technical users can ask these questions themselves through Spotter without writing SQL.

James’s team now focuses on expanding the semantic model and tackling higher-order analysis, rather than acting as a query translation service. The crowdsourcing model means subject-matter experts across the business can contribute definitions, with James’s team playing a validation and governance role.

Scenario: Priya is a data platform admin at a healthcare company. Her biggest anxiety is simple: what happens when an AI agent generates a query and returns data the user shouldn’t see?

With generic AI analytics tools, this is a real risk — the LLM generates SQL, and if the security rules aren’t embedded in the query itself, a user could receive rows they’re not authorized to access.

ThoughtSpot’s deterministic query generation applies row-level and column-level security as part of the query construction process — not as a post-processing filter. It’s architecturally impossible for a user to receive data outside their permission scope, regardless of how they phrase their question. Priya can roll out AI analytics to the entire organization with confidence, knowing governance is enforced at the Model layer, not just the UI layer.