microsoft/openaivec v2.5.0

v2.5.0

Repository: microsoft/openaivec

Tag: v2.5.0

Published: 2026-09-10T01:03:45Z

Prerelease: no

Release notes:

v2.5.0

Fabric Built-in Models on Spark Executors

Use Fabric runtime-managed authentication from Spark UDFs:

from openaivec import spark_ext

spark_ext.setup_fabric(spark)

• Supports string and structured Responses, Embeddings, prepared tasks, and parsing UDFs.
• Creates async Fabric HTTP clients inside Python workers and closes partition-owned clients after use. Driver tokens and live clients are not serialized to executors.
• Preserves existing batching, partition-local deduplication, input/output correspondence, and caller ownership of non-Fabric clients.
• Adds a runnable Spark notebook and a pinned Environment definition.

Import the Environment definition, use Full publishing, attach the Environment and a Lakehouse to the notebook, and start a new session. The verified configuration uses Fabric Runtime 1.3, Python 3.11.8, OpenAI SDK 2.0.0, gpt-5.1, and text-embedding-ada-002. Do not install openaivec[spark] over Fabric's managed Spark runtime.

See the authentication guide for setup and compatibility details.

Reliability and Packaging

• Includes SDK-side Pydantic JSON validation failures in schema inference's existing bounded retry loop, with validation feedback and the original retry limit preserved.
• Aligns the pandas minimum requirement to pandas>=2.3.3 across supported Python versions.
• Includes the public Fabric Spark notebook, Environment YAML, and authentication guide in the source distribution.

Validation

• CI: 505 tests passed; Ruff passed; Pyright reported no errors or warnings.
• The published 2.5.0 package was installed from PyPI through a dedicated Fabric Environment with external libraries and no custom wheel. The public notebook completed successfully on 2026-09-10, verifying all five UDF paths, six rows across two partitions, input/output correspondence, duplicate outputs, 1536-dimensional embeddings, and a repeated Spark action.
• The driver and both workers reported openaivec==2.5.0 and openai==2.0.0, with imported pandas 2.3.3 and PyArrow 19.0.1.
• Candidate 2.5.0.dev2 passed all five UDF paths on Fabric Runtime 1.3 with six rows, two partitions, two-row Arrow batches, duplicate inputs, 1536-dimensional embeddings, and repeated actions.
• The signed release wheel imports successfully in an isolated Python 3.10 environment. All 47 installed package files match the published wheel and merged source.
• Wheel and sdist SHA-256 values match PyPI's published metadata. Both Sigstore bundles verify against .github/workflows/publish.yml@refs/tags/v2.5.0 and the GitHub Actions OIDC issuer.

Compatibility Notes

Fabric built-in models are a preview feature with capacity, tenant, and regional prerequisites. Runtime 1.3 can expose older package metadata alongside newer imported libraries; successful UDF execution does not imply that every platform dependency warning is resolved. Long-running token refresh and scheduled execution identities have not been validated.

Changes: PR #180.

Documentation screenshots: Desktop | Mobile.