Catalyst query execution and recorded examples

Technical context for the recorded workflows. For the capability overview and narrated walkthrough, visit the Catalyst project page.The selected Catalyst contract uses a configured SQL source, its declared dialect, and every table and column the connection can read. A writer drafts SQL, a selected profile may add a reviewer, advisory findings remain visible, and the person chooses the exact query to run. Successful results can become Datasets, Widgets, and a Dashboard published to Superset.

Recorded locally with a writer and reviewer. Gemma 4 12B drafts and Qwen 2.5 14B reviews each query. These September 2026 recordings use a Spark source and show Superset rendering. These local recordings do not establish final server or owner acceptance.

Local product-flow recordings

Conversation turns

Model roles per turn

Configured source per session

Selected product contract

One source per sessionA session keeps one source identity, dialect, and readable schema. Choosing another source starts another session instead of mixing data or query context.

Complete readable schemaThe model, Available data view, and editor receive every readable table, view, column, and type. Optional descriptions may explain the data but cannot hide it.

Exact, person-controlled executionFormatting and validation are advisory. The person can edit the query, inspect findings, and run the exact selected SQL once to receive typed rows or the database error.

Product and deployment stay separateCatalyst is a SQL client. FHIR Data Pipes, Parquet, and Spark define the selected reference deployment without becoming a requirement for every Catalyst connection.

Watch the full product flow

Both walkthroughs show drafting alongside schema browsing, explicit execution, refinement, saved-SQL reuse, charts and Dashboard arrangement, followed by actual import and rendering in Superset. The pipeline introduction is 13 seconds; only labelled waiting periods are accelerated.

OpenELIS Laboratory

Local · writer and reviewer

3:04 · ~4 MB · recorded locally · aggregate results

Turn 1How many patients are there?

Turn 2Break down that patient count by gender, including patients with missing gender. Return gender and patient_count.

The second turn refines the exact current query rather than starting over.

OpenMRS HIV/ART program

Local · writer and reviewer

3:56 · ~4.7 MB · recorded locally · aggregate results

Turn 1How many CD4 count results were recorded in each month of 2026?

Turn 2Break down the monthly CD4 counts by gender, including missing gender.

The second turn refines the exact current query rather than starting over.

Go deeper

Catalyst application specification →

Try the public application. Open Catalyst to explore the deployed Workbench. The videos above were recorded in the local acceptance stack; dated evaluation findings are linked from the project evidence page.