How we built the semantic layer
Your semantic layer should learn from its mistakes.
AI has made querying data easier. It has also made something else much more important: teaching software what your business actually means.
A database can tell an AI system that a column is called status. It usually cannot tell it which statuses count toward pipeline, when an account should be excluded from a revenue calculation, why two apparently similar tables have different grains, or how your company defines an active customer.
That knowledge belongs in a semantic model.
But there is a catch. As semantic models become sophisticated enough to support AI agents, they become increasingly difficult to maintain manually.
That is why we built the Semantic Agent.
What is the Semantic Agent?
Supper's Semantic Agent is an AI agent for creating and maintaining the semantic model that sits underneath Supper's data agents.
It can investigate an identified semantic issue, help a user define a new business concept, examine relevant model context, and propose the changes required to make that knowledge reusable.
In normal interactive use, those changes remain under human control: the user approves what actually enters the semantic model.
The semantic layer has changed
Traditional semantic layers often center on a manageable set of metrics, dimensions, and relationships.
AI analytics demands considerably more context.
Supper's semantic model can include descriptions of schema components, table grain, enumerations, business Terms, Quirks in the data model, company-specific Calculations, technical instructions, reference code, and reusable Skills.
That richness is important because language models need more than column names to reason reliably about company data.
But every additional piece of context creates something that has to be authored, reviewed, corrected, and updated.
Eventually, maintaining the semantic layer starts becoming a data project of its own.
From fixing SQL to fixing the underlying knowledge
Consider a common scenario.
A data agent generates code. The code runs, but the result exposes a misunderstanding of the business.
The conventional response is often to fix the immediate output: modify the SQL, rebuild a view, change a dashboard, or document the exception somewhere.
That solves one instance of the problem.
Supper takes a different approach.
An Assessment can identify that the failure points to missing or incorrect semantic knowledge. The Semantic Agent can then begin with that context, investigate the relevant model components, and propose an update to the semantic model itself.
Fix the knowledge once, and future questions can benefit from the correction.
A conversation instead of a modeling project
The Semantic Agent can also be started without an Assessment.
A user might simply say, in effect: "There is an important concept in our business that Supper needs to understand."
The agent can investigate how that concept relates to the existing model, ask the user where business judgment is required, and propose a formal representation of it.
If several interpretations are plausible, it can present choices rather than silently making an assumption.
This changes the interface to semantic modeling.
Instead of asking every business user to think in SQL, schema definitions, or modeling syntax, Supper can meet them in conversation while still producing technical semantic-model components underneath.
For some workflows discussed during development, work that previously involved hours or even days of SQL and model maintenance can become a conversation measured in seconds.
Human-controlled when it matters, autonomous when it makes sense
An agent editing the layer that defines your business raises an obvious question: who gets the final say?
In the standard interactive workflow, the user does.
The Semantic Agent proposes changes. The user can review, choose among options where necessary, and approve or reject the result.
There is also a different mode during onboarding.
When Supper is creating the semantic model for the first time, the Semantic Agent can run headlessly and generate foundational components automatically. Those components give humans a much stronger starting point to review than a blank semantic model.
The distinction matters. Autonomy is useful where the goal is acceleration. Human judgment remains important where company-specific meaning is ambiguous.
The larger idea: an agentic semantic layer
The most interesting thing about the Semantic Agent is not the chat experience.
It is the feedback loop.
AI agents depend on semantic knowledge to produce good data work. Their actual data work also reveals weaknesses in that semantic knowledge.
Once those weaknesses can be detected and fed back into an agent that improves the model, the semantic layer stops behaving like static configuration.
It can evolve through use.
That is the direction we think semantic systems need to go: not simply richer semantic layers, but agentic semantic layers that help build, maintain, and improve the context AI depends on.
Because the more important your semantic model becomes, the less reasonable it is to maintain all of it by hand.
FAQ
What is a Semantic Agent?
A Semantic Agent is an AI system that helps create and maintain the business and schema knowledge used by data agents to interpret company data.
Does Supper automatically change business definitions?
In normal interactive use, the Semantic Agent proposes changes and the user has final approval. During initial onboarding, it can also perform foundational semantic-model work headlessly for later review.
What kinds of things can the Semantic Agent help with?
Use cases established during development include business Terms, model Quirks, technical descriptions, reference code, semantic issues surfaced through Assessments, and new concepts introduced directly by users.
Why is this better than correcting a query manually?
A query fix addresses one output. A semantic-model correction can improve the reusable knowledge available to future questions.