Your agent does not know which of forty tables is authoritative
Text-to-SQL against a bank’s schema does not fail loudly. It returns a plausible number from the wrong table, using the wrong definition of exposure, and nothing errors. TopQuadrant holds the governed definitions, permitted relationships, and entitlements that turn an agent’s question into a query you would defend in front of a model risk committee.
Trusted to govern the data behind the world's most regulated decisions.
The semantic layer is where the argument about the number gets settled
Finance, risk, and the front office each define exposure, revenue, and active customer differently, and each definition has been copied into a dbt model, a dashboard formula, and someone’s notebook. Agents inherit all of it. Putting the definition in one governed place, mapped explicitly to physical data, is the only version of this that scales past a pilot.
More accurate AI answers when the model reasons over governed context rather than raw schemas.
Cheaper to operate than rebuilding context per application and per agent.
One governed definition per metric, resolved at query time across BI, applications, and agents.
Governed definitions
Settle the definition once, in a place with an owner
Every metric and entity carries its business definition, its owner, its calculation, and its mapping to physical columns. Where trading, risk, and finance legitimately need different views, the difference is modelled explicitly rather than left to diverge.
- Owner per definition. Disagreement resolves to a named person and a governed change process.
- Divergence modelled. Where two functions need different logic, both are governed and related.
- FIBO-aligned. Import the Financial Industry Business Ontology and extend it with your own model.
Constrained query construction
Remove the guessing from query generation
Agents receive concepts, permitted relationships, and metric logic rather than a schema dump. The join paths available to them come from the ontology, so the class of failure where a model invents a plausible join is not available.
- Ontology-bounded joins. Only modelled relationships are traversable.
- Definitions before generation. The agent resolves what a term means before composing a query.
- Open interfaces. Governed context served over API, SPARQL, and MCP for any agent framework.
Entitlements for agents
An agent sees what its caller is entitled to see
Access rules attach to governed concepts and are evaluated per request against the identity the agent is acting for. A relationship manager’s assistant and a risk analyst’s assistant read the same governed layer and get different data, correctly.
- Per-request evaluation. Entitlements resolve at query time, not at index build.
- One policy model. The same rules govern dashboards, notebooks, and agents.
- Prompt-resistant. Access is enforced in the data layer, so phrasing does not change the outcome.
Model risk evidence
Produce the trail your validation function asks for
Under revised interagency model risk guidance, an LLM used in a credit or AML workflow is in scope for validation, documentation, and monitoring. Every served answer records the definitions resolved, their versions, the entitlements applied, and the data returned.
- Answer-level records. Question, resolution path, versions, and result retained and queryable.
- Reproducible. Re-run a historical question against the definitions in force that day.
- Inventory as data. Agents, owners, data domains, and validation status held in the same graph.