SQLMesh

pip install "dataveil[sqlmesh]"

from sqlmesh.core.context import Context
from dataveil.adapters.sqlmesh import SQLMeshAdapter

adapter = SQLMeshAdapter(Context(paths="path/to/project"))

Reads go through the project's virtual layer, so a model is addressed by its plain schema.model name. Writes go to the model's current physical snapshot table, because a virtual-layer view is generally not updatable.

On DuckDB and Postgres projects the adapter registers the checksum functions, so every classifier works. On DuckDB projects it also registers the person-name function. On other engines, the credit-card, IBAN and name classifiers are skipped, and the regex classifiers work.

Cleansing is not durable yet

An operation changes the physical snapshot table outside SQLMesh's own state. The next sqlmesh run or plan of a FULL model rebuilds the table and undoes it. Durable cleansing, written into the model's query, is planned for 0.5.

Used by sqlmesh-mcp's profile_model, propose_cleansing_plan and apply_cleansing_plan tools. With dataveil's own MCP server: DATAVEIL_ADAPTER=sqlmesh and SQLMESH_PROJECT_PATH.