Changelog

All notable changes to TabuLLM are documented here.

[1.4.0]

• Added a Sphinx documentation site (Furo theme, MyST Markdown, autodoc/autosummary with Napoleon), published to GitHub Pages at https://asmahani.github.io/TabuLLM/. Covers the README as a landing page, full API reference for tabullm.__all__, an examples page linking the two tutorial notebooks with their external requirements, and a paper/reproducibility page. A new docs optional dependency group and a docs.yml CI workflow build the site (warnings as errors) on every PR/push and deploy it on pushes to main.

• Fixed several stale and incorrect examples in public docstrings, found while building the documentation: ClusterExplainer’s class docstring called a nonexistent GMMFeatureExtractor.fit_predict() and the obsolete count_tokens_only=True argument (now preview=True); GMMFeatureExtractor’s example passed an unsupported text_columns= argument to TextColumnTransformer; several examples referenced undefined names (model, RandomForestClassifier, df, emb_model, chat_model) without importing or defining them. Examples are now self-contained and runnable (credentialed calls are noted in comments).

• Added the Documentation link to [project.urls], a documentation badge to the README, and set the GitHub repository homepage to the published docs site.

• Widened the supported langchain-core range to >=1.3,<1.6 (from <1.5); CI now tests both the minimum supported line and the latest allowed release explicitly.

• Added the SoftwareX fraud-benchmark reproduction package (paper_experiments/): pinned environment, run/summarise/plot scripts, and archived per-fold results, metrics, and rolled-up summaries. Added the published SoftwareX paper citation and DOI to the README.

• Excluded paper_experiments/paper-requirements.txt (a pinned snapshot of the environment used for the paper’s benchmark runs) from Dependabot, since it is not meant to track upstream releases.

[1.3.0]

• Added README badges for CI, coverage, PyPI version, supported Python versions, and license, and enabled Codecov uploads from the test workflow.

• Reduced the core runtime dependency surface to langchain-core; direct langchain and langchain-community requirements were dropped because TabuLLM uses core abstractions plus provider-specific integration packages.

• Tightened the supported runtime window for the LangChain stack to the tested langchain-core 1.3.x-1.4.x series, with matching caps on numpy, scipy, scikit-learn, and pydantic.

• Expanded CI to run the test suite across multiple LangChain minor lines and updated tests to use langchain_core.embeddings.Embeddings directly instead of the legacy compatibility shim.

• Added weekly Dependabot checks for Python dependencies and GitHub Actions so upstream changes surface as reviewable PRs before they reach users.

• Breaking: the minimum supported numpy version is now 1.26, the minimum supported pydantic version is now 2.7, and the package no longer installs the umbrella langchain / langchain-community packages as transitive runtime dependencies.

[1.2.1]

• Fixed deprecated LangChain import in embed.py (langchain.embeddings.base → langchain_core.embeddings).

• Removed phantom pillow runtime dependency.

[1.2.0]

• Removed SphericalKMeans class. For L2-normalized embeddings, sklearn’s KMeans is mathematically equivalent; GMMFeatureExtractor provides strictly richer features for pipeline use.

[1.1.0]

• Added multiple testing correction to explain() via the correction parameter ('bonferroni', 'holm', 'fdr_bh'). When set, a P-value (adjusted) column is appended to the per-cluster results and, when observation_stats is provided, to the stat-association table. Backward compatible: default is None (no correction).

[1.0.3]

• Fixed broken package installation (1.0.2 wheel was published without Python source files).

[1.0.2]

• Fixed __version__ mismatch; aligned __init__.py with pyproject.toml.

[1.0.1]

• Switched fraud dataset download from Kaggle to Zenodo (no credentials required).

[1.0.0]

• Initial release.