Release 0.14.1

• Fix incompatibility between intel-mkl-tool and anyhow > 1.0.48.

Release 0.14.0

• Update finalfusion dependency to 0.17, this brings many improvements, including:
  • Support for floret embeddings.
  • Improved errors, providing structured context.
• Relicense to Apache License version 2.0 or MIT

Release 0.13.2

• Use cross-building to build for more platforms.

Release 0.13.1

• Fix linkage when the intel-mkl-amd feature is enabled.
• Use the LP64 interface, rather than the ILP64 interface, this fixes some functionality that failed with MKL.

Release 0.13.0

• Support the Accelerate framework on macOS for BLAS/LAPACK.
• Improved support for Intel MKL.
• Update to finalfusion 0.16, reductive 0.7, and ndarray 0.15.

Add the bucket-to-explicit/select subcommands, support angular similarity

• Add bucket-to-explicit. This subcommand converts a finalfusion/fastText embedding file with a bucket-based n-gram vocabulary to an explicit n-gram vocabulary. (by @sebpuetz)
• Add an --similarity option to the similar/analogy subcommands. This option allows selecting the similarity measure. Cosine similarity is still the default, but angular can be used to switch to angular similarity. (by @danieldk)
• Add select subcommand. This subcommand creates a new embedding matrix from an existing embedding matrix that covers the given set of words. (by @danieldk)

Always link against system BLAS

Always link against system OpenBLAS/MKL/Netlib BLAS

Licensing

Provide Apache License version 2.0 as an additional licensing option.

Release 0.11.0

• Add reconstruct subcommand to reconstruct quantized embeddings. (@RealNicolasBourbaki and @danieldk)
• Better error handling and reporting through the anyhow crate. (@danieldk)
• Compile ndarray with BLAS support when one of the openblas, intel-mkl, or netlib features is enabled. (@danieldk)
• Build finalfusion-utils directly from the repository with Nix. E.g.,

  $ nix-env -i -f https://github.com/finalfusion/finalfusion-utils/tarball/0.11.0

  Installing through the Nix package set is still recommended for faster builds. (@danieldk)
• BucketIndexer optimizations through finalfusion 0.12. (@sebpuetz)

Update to finalfusion 0.11.0

This release updates the finalfusion dependency to 0.11.0. This adds support for explicit n-gram vocabularies, which are introduced in finalfrontier 0.7.0.