Cassis Ontology Skills: An Open-Source Alternative to Nvidia Auto-Ontology
Automate warehouse documentation and metric definitions using evidence-backed, local-first LLM workflows.
1Automating Warehouse Context
Cassis provides a framework to build and expand data ontologies by extracting evidence from existing assets. Unlike black-box solutions, it treats documentation like code, allowing users to review generated metrics and definitions. It connects directly to dbt projects and schema exports to ensure that every description is grounded in actual SQL logic.
2Evidence-Based Documentation
The tool avoids hallucinated column descriptions by requiring concrete evidence from the warehouse. It processes dbt projects, schema exports, and existing documentation to map meanings accurately. On the GitLab public analytics project, it successfully processed 2,573 out of 2,641 columns, leaving only 68 blank rather than guessing at potentially incorrect definitions.
3Local-First Execution
Cassis runs entirely on your local machine, ensuring data privacy and security. No telemetry, API keys, or account connections are required for the core processing stages. The only data leaving the laptop consists of specific queries sent to the LLM during the judgment phase, keeping your warehouse metadata and sensitive structures private.
4Deterministic Workflow Stages
The process is divided into four distinct checkpoints, allowing users to review outputs at every stage. Instead of a single opaque output, the tool generates a file tree that includes domains, tables, metrics, and joins. This structure ensures that rules, such as mandatory filters for revenue calculations, are carried end-to-end.
5Handling Ambiguous Questions
When the system encounters questions it cannot resolve, it logs them in an open-questions file rather than making assumptions. This prevents the propagation of incorrect data definitions. By separating settled facts from unresolved queries, the tool ensures that users only accept documentation that has been verified against the underlying SQL evidence.
6Integration with dbt
The kit integrates seamlessly with dbt projects, allowing users to merge generated ontologies back into existing schema.yml files. It supports various SQL dialects including BigQuery, Snowflake, and DuckDB. By utilizing dbt manifests, it captures complex relationships and grain definitions that are often lost in manual documentation efforts or generic schema exports.
7Metric Corroboration Mechanism
Cassis validates metrics by comparing expression shapes against the warehouse SQL. If a metric formula is defined, the tool ensures that mandatory filters, such as status flags, are included. This prevents common errors where revenue is miscalculated due to missing filters, providing a transparent audit trail for every metric in the system.
8Scalable Incremental Growth
The ontology is built incrementally, typically in batches of 20 to 30 tables per run. This approach keeps costs predictable and manageable, regardless of the total warehouse size. Users can focus on specific domains first, ensuring high-quality output without the need to process the entire data warehouse in one expensive session.
9Advanced Judge Pass
A critical component of the run is the judge pass, an agent that re-reads every drafted claim against the provided evidence. This stage has proven effective in catching descriptions that contradict the actual SQL logic. By automating this verification, the system ensures that the documentation remains consistent with the physical database implementation.
10Flexible Input Sources
While dbt is the primary supported input, the tool accommodates other sources like SQLMesh or generic schema exports. It can ingest free-form documentation, wikis, and PDFs to answer specific blocking questions. This multi-modal approach allows the system to synthesize information from various organizational silos into a single, coherent, and machine-readable ontology.
11Operational Transparency
The tool provides clear reports in OUTPUT.md, mapping the entire ontology structure. Users can verify the results by reviewing the generated YAML files, which contain grain definitions, column descriptions, and data types. This transparency allows data teams to maintain full control over their semantic layer while benefiting from automated documentation generation.
12Strategic Data Governance
Cassis serves as a robust alternative for teams needing to bootstrap context on complex data projects. By prioritizing evidence over guesswork, it reduces the risk of data misinterpretation. It transforms documentation from a tedious manual task into a verifiable, automated process that scales alongside the growth of the organization's data infrastructure.