Databricks gives Genie Code file context—but keeps uploads transient by default

Azure Databricks added file uploads to Genie Code on August 14, giving its coding and data assistant a second path to context alongside the tables, columns, descriptions and lineage it already reads through Unity Catalog. The new path is deliberately temporary: attached files apply only to the current chat and are not saved to the workspace file system unless the user explicitly asks Genie Code to save them. Microsoft’s August release notes and updated usage guide document the change.

What the upload surface covers

Users can attach files from the prompt box or drag them into a chat. Supported formats include PNG, JPEG, GIF and WebP images; PDF documents; CSV, XLSX and XLS spreadsheets; and JSON or JSONL data. That makes it possible to give the agent context that has not first been modeled as a governed table—for example, a workbook, an exported result set or a PDF specification—while keeping the file scoped to one conversation by default. Microsoft’s guide lists the formats and persistence behavior.

Genie Code is broader than a text-to-SQL interface. Databricks describes it as an agent for generating and running code, building pipelines and dashboards, debugging errors, and working with Unity Catalog assets across notebooks, the SQL editor, Lakeflow, dashboards and MLflow. It can also receive table names, column descriptions, samples and cell outputs as model context, subject to the user’s Unity Catalog permissions. The product overview and usage documentation define that scope.

The boundary matters

File context does not become governed workspace data automatically. That distinction limits accidental persistence, but it also means an uploaded workbook or PDF is not, by virtue of attachment, a reusable semantic definition for other chats or users. Persistence requires a separate save request, while durable organizational guidance belongs in workspace instructions, skills or governed data assets. Microsoft’s documentation exposes those as separate controls.

The release therefore expands what Genie Code can inspect without collapsing temporary evidence into the governed data layer. For NL2SQL teams, the useful question is no longer only whether an agent can see a schema; it is also whether the context it receives is transient chat material or a managed asset that can be reused and audited.