[vishal-bala]

Motivation

The jobs that run on a single Python version use 3.11, which is no longer where most redisvl installs are. 3.12 now holds the largest share of downloads and is not close to end of life, while 3.11 has fallen to third. The figures below are from pypistats.org, with mirrors excluded, checked on 2026-11-02:

Python	Last 30 days	Last 208 days
3.12	44.1%	46.2%
3.13	22.2%	19.0%
3.11	18.8%	20.1%
3.14	9.1%	not checked
3.10	1.0%	not checked

Each column is a share of all downloads in that window, and versions not listed make up the remainder. 3.11's share is lower over the last 30 days than over 208, so it is shrinking.

The service tests and notebook validation are the only CI runs that call live provider APIs, so they should run on the version most users have.

Changes

Default Python version moves from 3.11 to 3.12

Each workflow sets one PYTHON_VERSION for its single-version jobs. This PR changes that value, and the Read the Docs interpreter, from 3.11 to 3.12:

• .github/workflows/test.yml: the service tests (make test-all, which calls external provider APIs), notebook validation and the docs build job.
• .github/workflows/test-fork-pr.yml: the maintainer-triggered service tests for fork PRs.
• .github/workflows/auto-release.yml: the canary build, the version bump step (uv lock --check) and the release build. The wheel is pure Python (py3-none-any), so this one only keeps the workflows consistent.
• .readthedocs.yaml: build.tools.python.

The test matrix in test.yml and the style-check matrix in lint.yml still cover every supported version, 3.11 to 3.15, so no version loses coverage.

Notes

This PR is stacked on top of the Python version stack (#768, #810, #811, #812) rather than branched from main, so it lands after the 3.10 drop and the version-range changes it sits with. It does not depend on any code in those PRs: the diff against #812 is the four lines above.

The Makefile's black --target-version py311 stays as it is. That flag is not a runtime default: it tells black which Python grammars its output must stay valid for, and it overrides the target-version list in pyproject.toml. With 3.10 dropped, py311 is exactly the supported minimum, and raising it to py312 would let black emit syntax that 3.11 cannot parse. On the current tree black 26.5.1 produces identical output for py311 and py312.

Reviewer Brief

Start here: .github/workflows/test.yml, env.PYTHON_VERSION. It drives the service tests and notebook validation, the only CI runs that call live provider APIs, as well as the docs job.

Where it could fail

• Service tests on 3.12: a provider SDK pinned in uv.lock that misbehaves on 3.12 only when it calls its live API would surface here first, since the matrix legs on 3.12 skip the API tests.

Not verified: the service tests and the API-dependent notebook cells, because no provider keys were available locally; the Read the Docs build on Read the Docs itself; the release workflow, which runs only after merge.

Beyond CI: on Python 3.12.12, tests.test_imports passes with only the base dependencies installed, and uv build succeeds (the wheel is py3-none-any); uv lock --check passes under the pinned uv 0.13.1. The docs build in a uv sync --group docs environment gives the same 12 Sphinx warnings on 3.11.9 and 3.12.12. The notebooks against a local redis:8.4 give identical per-cell results on 3.11.9 and 3.12.12. Fifteen cells in 03_llmcache, 04_vectorizers and 06_rerankers fail on both for want of provider API keys; six !rvl CLI cells hit the 60-second cell timeout on both while the two runs shared the machine, and pass when their six notebooks are re-run alone on 3.12.

_{Reviewed ad90047 · Claude Code (claude-opus-5-5) · lenses: correctness, security, design, tests, docs, Redis behaviour}