How We Use Genie AI at Yoma Fleet
At Yoma Fleet, we leverage Databricks AI/BI Genie to democratize data access and support our data‑driven decision-making without forcing every user to learn SQL. Within Databricks, Genie acts as a conversational SQL interface that transforms natural language questions into governed SQL executed against our Delta Lake, respecting Unity Catalog governance such as row‑level security and column‑level masking. Business users receive answers or visualizations directly, removing the need to build and maintain dozens of isolated BI dashboards.
We currently maintain two primary Genie spaces for two products. Each space is crafted around a carefully curated semantic data layer and example questions that teach the model how we talk about our business. This structure ensures that Genie both understands our domain vocabulary and complies with our security and governance policies.
What is Genie?
In our context, Genie is Databricks AI/BI’s conversational analytics engine that helps users ask plain‑English questions about data. Behind the scenes, Genie interprets natural language and generates SQL that runs on a SQL Warehouse against curated tables and views registered in Unity Catalog. It isn’t a mystery “AGI” tool, and it isn’t a legal tech contract assistant just prompt‑to‑SQL with strong governance backed by metadata, curated examples, and structured instructions created by analysts.
Supported Data Domains and Questions
Our Genie spaces focus on specific business functions, with domains and example questions tailored to each:
| Space | Main Data Sources | Questions We Answer | Key Metrics |
| Product 1 | Sales, transactions, branch performance, org status | “Daily sales count for Jan 2025?” “Which branches drive most revenue?” | Sales volumes, DAU/WAU, SLA metrics |
| Product 2 | Reservations, installs, ratings, cancellations | “How do ratings vary by vehicle type?” “Which periods saw spikes in installs?” | Booking trends, client usage stats |
By guiding Genie with curated metadata, synonyms (e.g., “branch” ≈ “store”), and example questions, we’ve driven higher accuracy in SQL generation and faster results for business users.
Architecture and Data Modeling
In the Yoma Fleet engineering stack, our Delta Lake follows a classic layering model:
- Bronze: Raw ingested records, minimal cleaning.
- Silver: Normalized and consistently typed data.
- Gold: Curated business views and aggregates optimized for analytics.
We expose only curated views in Unity Catalog to Genie spaces: raw tables are never directly visible. Views are designed with explicit column descriptions, synonyms, and semantic clarity so that Genie can map natural language prompts to correct SQL queries. This modeling discipline is critical because Genie leans on metadata and curated instructions to interpret user intent correctly.
Every Genie space is effectively a domain context that includes:
- A defined catalog and schema with permissions.
- Starter questions and synonyms aligned with business language.
- Example SQL patterns and curated semantic metadata.
This context helps Genie generate consistent and governed SQL that respects masks and row‑level access, reducing the risk of accidental data exposure.
Governance, Security, and Observability
Genie’s integration with Unity Catalog allows us to enforce data governance at scale. Row‑level security ensures that users only see the data they’re entitled to, and column‑level masking protects sensitive fields such as PII. Query execution limits and monitoring help us control costs and performance, while audit logs give insight into usage patterns.
We observe the health of each Genie space via metrics like query success rate, execution latency, cache hit rate, and classification of errors due to schema changes or ambiguous prompts. This operational visibility lets us refine both the semantic layer and example sets to improve deliverability over time.
The Prompt‑to‑SQL Reality
The true power of Genie lies in SQL generation that you can inspect and reuse. For example, a natural‑language prompt like “Daily sales in January 2025” is translated into a fully executable SQL query that runs against our gold sales views. This transparency means developers and engineers can validate, optimize, and reuse SQL logic as needed.
Making Genie Accurate (and Efficient)
Engineering discipline plays a central role in keeping Genie both accurate and cost‑efficient. We invest in:
- Semantic naming conventions: clear, intuitive table and column names.
- Complete metadata: descriptions, synonyms, and usage context.
- Curated views: pre‑joined Gold views with unambiguous granularity.
- Evaluation harnesses: automated testing of canonical questions to validate outputs, catch regressions, and maintain trust in queries over time.
When adding a new metric, our engineers start by creating or updating a curated view with clear metadata and a defined grain. We enrich the view with synonyms and example questions in Genie space configuration, then write automated tests to check both shape and logic against known samples. Once validated, it is deployed and exposed to business users via the configured Genie space.
Issues that arise from schema changes or ambiguous prompts are resolved by either aliasing old column names or enriching the Genie space with new synonyms and examples.
Real Business Impact
The introduction of Genie has shifted analyst effort from writing repetitive SQL boilerplate to focusing on domain questions and model accuracy. Business stakeholders now self‑serve common queries, while engineers maintain a documented semantic layer rather than a forest of ad‑hoc dashboards.
Is it perfect? Not yet. Genie still depends on clean schema surfaces and good metadata quality. But when aligned with robust engineering practices, conversational analytics becomes a productive collaboration between business and data engineering teams.
