Skip to content
RelationalAI Docs

Python package and CLI 1.32.0

If you use prescriptive reasoning (Public Preview, by request), review optimization problems after upgrading because some formerly accepted constraints and objectives now error or change meaning.

How to install this version

Review the changes below before installing. To install this version, activate a compatible Python virtual environment and run:

Terminal window
python -m pip install "relationalai==1.32.0"

For the latest installation instructions, see Start building with PyRel.

Breaking Changes

Section titled “Breaking Changes”
  • Prescriptive reasoning (Public Preview, by request): When you create an optimization problem in your Python code, some constraints and objectives that used to be accepted now raise a clear error or change meaning. Review your existing problems after you upgrade. For example, applying count directly to a decision variable is now rejected because it counts rows rather than the values the solver chooses. A few others still build but now describe a different problem. To check them, print the problem with Problem.display() and confirm the formalized constraints and objectives match what you intended.

New Features and Enhancements

Section titled “New Features and Enhancements”

  • You can now use the percentile and median aggregates to summarize your data. You can group either one with per() like any other aggregate.

  • Snowflake object tags: When you deploy a model to Snowflake, you can now apply your own existing Snowflake object tags to every object the deploy creates by listing them under deployment.tags. On a redeploy, each object keeps only the tags you list.

    deployment:
      tags:
        GOVERNANCE.TAGS.COST_CENTER: engineering

  • Row-access policies: You can now attach an existing Snowflake row-access policy to the tables, views, and dynamic tables a deploy creates by adding a per-output policies override under deployment.outputs. RelationalAI does not fully check the policy first, so an invalid one can still fail when the deploy runs.

    deployment:
      outputs:
        overrides:
          - objects: [ORDERS]
            policies:
              row_access_policy:
                fqn: GOVERNANCE.POLICIES.TENANT_ACCESS
                columns: [TENANT_ID]

  • Existing ownership tag: You can now point RelationalAI at a Snowflake tag you already created for object ownership with deployment.managed_tag. A governed account can then deploy a model without the Snowflake CREATE TAG privilege. The deploying role needs APPLY on the tag and USAGE on the database and schema that hold it.

    deployment:
      managed_tag: GOVERNANCE.TAGS.RELATIONALAI_MANAGED

  • You can now open a rai doctor report in the rai debugger offline, without access to the Snowflake account it came from. The trace, Refresh Log, and deployment graphs all work from the unzipped report.

Bug Fixes

Section titled “Bug Fixes”

  • A model with a wide concept, roughly 43 or more nullable properties, that failed to deploy with a RecursionError now deploys again. This fixes a regression introduced after version 1.29.1.

  • Fixed an issue that caused some models to fail to deploy with an “object already exists” error.

  • Custom metadata schema: If you place a model’s metadata and operation log in a custom schema with deployment.meta_schema, RelationalAI now uses that location consistently for deploys, teardown, and branch operations. Changing the setting later does not move existing metadata, so set it before your first deploy.

  • Redeployed objects: Redeploying a model now recreates every changed Snowflake object consistently, and it preserves the grants you added to its procedures and functions. Because objects are recreated rather than altered, you may need to restore other object-level state yourself.

  • Deploying a model to Snowflake is faster, because RelationalAI now runs its two metadata setup phases in parallel.

  • Conflicting deployment records: Two deployment records that claim the same model identity no longer break rai models listing or teardown for the whole database. The conflicted model stays listed with a warning that explains how to repair it.

  • For models deployed to Snowflake, the variance and variance_samp aggregates no longer lose precision on whole-number input. Expressions that raise a value to a power with **, such as math.haversine, no longer fail. This does not fix geometric_mean over negative input on Snowflake, which is still unsupported.

  • For models deployed to Snowflake, comparing two entity types with == no longer errors when their identities are stored differently, for example a generated hash versus an integer identify_by key. Because the two kinds of identity can never be equal, the comparison now returns False.

  • A where() filter whose condition is a Python builtin call with a literal argument, such as abs(-1) == 2, now filters when the condition is false, as expected.

  • For models deployed to Snowflake, queries that compute several aggregate measures over the same grouping now run in a single pass instead of re-scanning the data once per measure. The results are unchanged.

  • Calling inspect() or to_df() on a define or require fragment now raises a clear error, “Cannot inspect a define” instead of crashing. A define or require adds to your model rather than running as a query, so there is nothing to inspect.

  • Configuration load errors: When a configuration file that RelationalAI discovers automatically, such as ~/.rai, the Snowflake CLI config, or a dbt config, fails to load, RelationalAI now reports the real error and names the file that failed, instead of a generic “no configuration found” message. A commented-out {{ env_var(...) }} line is also no longer evaluated, so it can no longer fail on an unset variable.