[gouyt13]

Summary

Prepare the 0.8.0 release.

• Bump the version to 0.8.0 in pyproject.toml and CMakeLists.txt.
• Update install pins and version text in the README, quick start, installation guide, C++ integration guide, docs home page and the Python sample README.
• Add the 0.8.0 release notes (docs/docs/releases/0.8.0.md), their navigation entry and a README news entry. The notes cover the train/add/search API (feat(index): train centroids inside IVF and HNSW indexes #203, feat(python): train, set_centroids and add for IvfIndex and HnswIndex #204, feat(python): train, add and search for SymqgIndex #205), its deprecations and behavior changes, k-means sampling and assign (feat(clustering): add optional k-means sampling and assignment of new vectors #202), the breaking C++ changes (the g_ accessor rename from refactor!: rename query factor accessors with the g_ prefix #201, the removed rabitqlib::Array header from refactor: simplify core and utils and remove unused Array #175, and saving empty indexes), seeded-rotation and t_const changes off GCC 11+ (feat(core): portable seeded rotations and a frozen fast-mode scale table #199), and the validation, ISA-isolation, allocation and atomic-save fixes.
• Move benchmarks/run.py off the deprecated forms. It still clusters with RaBitQKMeans itself and passes the result through set_centroids and add(..., cluster_ids=...), so the benchmark protocol and its recorded configuration (quantization_bits) are unchanged.

Testing

• python -m mkdocs build --strict --config-file docs/mkdocs.yml, git diff --check, ./scripts/check-python.sh
• ruff format --check .github/scripts, ruff check .github/scripts, python -m unittest discover -s .github/scripts -p 'test_*.py' (60 tests)
• benchmarks/run.py on 3,000 synthetic 64-dim vectors with PYTHONWARNINGS=error::DeprecationWarning: rabitq-ivf, rabitq-hnsw, rabitq-symqg (--quantization-bits 0 and 4) all complete. With --threads 1, the old and new runner write byte-identical HNSW index files. This is a smoke test, not a performance measurement.

Downstream compatibility

Platform: Intel Xeon Gold 6418H, Ubuntu 22.04.5 (kernel 6.8.0), GCC 14.3, CMake 3.28.3. The txtai candidate is a package built from this branch's working tree with RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF. The zvec candidate is a git archive of 08bc325, this branch's base, whose library sources differ from this branch only in the version in CMakeLists.txt.

txtai master at 2a11bb1cfb (its RaBitQ backend last changed in 4d5e05f), Python 3.13, run in a temporary virtual environment:

python -m pytest test/python/testann/testdense/testrabitq.py

rabitqlib	Result
0.7.2 (PyPI)	5 passed, 4 subtests passed
0.8.0 candidate	5 passed, 4 subtests passed; 74 warnings, all DeprecationWarning from txtai's use of the old IvfIndex/HnswIndex constructors and build

This covers the backend's IVF and HNSW build, query, save and load, and IVF append and delete. txtai keeps working in 0.8 but will break in 0.9 unless it moves to the 0.8 API.

zvec main at bc7a90b258, which pins RaBitQ-Library v0.7.0 (de793e8) as a submodule and builds core and FastScan only (RABITQ_BUILD_INDEX=OFF, OpenMP off). Configured with -DCMAKE_BUILD_TYPE=Release -DBUILD_PYTHON_BINDINGS=OFF -DBUILD_TOOLS=OFF -DBUILD_C_BINDINGS=OFF; RABITQ_SOURCE_DIR pointed at the candidate archive.

Test	v0.7.0	0.8.0 candidate
ivf_rabitq_streamer_test	24 passed	24 passed
hnsw_rabitq_streamer_test	9 passed	9 passed
rabitq_dispatch_test	1 passed	1 passed
hnsw_rabitq_index_test	4 passed	4 passed

The candidate requires one change in zvec: tests/core/algorithm/ivf_rabitq/ivf_rabitq_streamer_test.cc calls kbxsumq(), which #201 renamed to g_kbxsumq(). Without that change the test fails to compile (no member named 'kbxsumq'; did you mean 'g_kbxsumq'?); with it, all tests above pass. zvec's library sources need no change.

Not tested: other zvec targets and its Python bindings, txtai outside its RaBitQ test module, platforms other than Linux x86-64, and native-optimized builds downstream.

Performance

No performance measurement was made for this release. The k-means sampling benchmark was not run, and the 2,000-cluster threshold for QGKMeans in index train is unmeasured; the release notes say so. The published benchmarks keep their 0.7.1 provenance.

Merging this into main triggers the automatic release: once Test and Python Wheel succeed for the merge commit, Release wheels tags v0.8.0, publishes to PyPI and creates the GitHub Release.