Release v1.1.0

Changed

• Remove dependency on the protobuf package by using a Rust implementation for serialization

Fixed

• Tests were failing because of a breaking change in Nox

Release v1.0.1

Changed

• Update pre-commit hooks and dependencies
• Allow to use also protobuf 4

Release v1.0.0

Added

• The storage backends do have now a method query_documents() to leverage economies of scale when
  querying multiple documents at once.

Changed

• The public interface of the library is declared stable, hence it is ready for version 1.0.
• char_ngrams() is now fully implemented in Rust, giving a speedup of 2x.
• minhash LSH uses the new query_documents() of the storage backends instead of running concurrent
  queries.

Fixed

• Wrong operator precedence in minhash implementation, which lead to incorrect results.
• Incorrect parsing of tokenize argument for SimilarityStore for char_ngrams without padding.

Release v0.10.0

Changed

• Improved performance of SimilarityStore.query_top_n()

Release v0.9.3

Fixed

• Fixes #63 which led to Exceptions in case of empty documents.

Release v0.9.2

Fixed

• Fixes #62 which led to TypeErrors in case of multiple identical results.

Release v0.9.1

Changed

• Minimum number of hash permutations for Minhash LSH set to 16 to avoid artifacts as described
  in #61.

Release v0.9.0

Added

• ScyllaDBStore now accepts a table_prefix setting.

Changed

• The classes in narrow_down.data_types were moved to narrow_down.storage.
• The initialize() method of the storage backends can now be called multiple times without issues.

Fixed

• A use of collections.Counter as typehint broke mypy checks.

Release v0.8.0

Added

• Direct InMemoryStore file serialization in the Rust backend.
  This avoids a memory peak and also improves the performance of the operation compared to
  (de-)serialization via the detour of a Python bytes object.

Release v0.7.0

Added

• InMemoryStore can be serialized to and deserialized from MessagePack.
• SimilarityStore.top_n_query() now allows to find a limited number of most similar documents.
• SimilarityStore offers the option to validate the similarity score if the document is available
  to avoid false positives.

Changed

• SimilarityStore objects can now be created by a factory coroutine create() instead of
  calling first __init__() and then initialize(). This makes the usage of the class more
  straight-forward.
• The exact_part of a document is now also stored in storage level "Document".
• The InMemoryStore no longer uses Python dictionaries as storage, but rather a class in the Rust
  extension to reduce the memory footprint by a lot.

Fixed

• The number of partitions is now stored in the database for the SQLite backend. This way the DB
  is self-contained and the user doesn't have to keep the number elsewhere.