All notable changes to charset-normalizer will be documented in this file. This project adheres to Semantic Versioning. The format is based on Keep a Changelog.

• Raised the Cython upper bound to <3.4 for native builds. The bound remains <3.3 for abi3 builds to preserve compatibility with the Python 3.7 Limited API.

• Valid UTF-8 Chinese JSON incorrectly detected as PTCP154 due to excessive noise penalties for uncommon CJK characters. (#796)
• Supported encodings without aliases failing name resolution or being ignored in charset declarations. (#800)

• Raised upper bound of setuptools to v84 (#794)
• Cache performance access optimization for our CharInfo struct (prebuilt only).

• No longer decoding large content when the noise detector output give a high entropy. Only impacted large content input >1M bytes.

• Explicit support for Python 3.15

• Comparing a CharsetMatch to a non-alias encoding strings (#773)
• Return 0.0 CharsetMatch.multi_byte_usage for empty payloads instead of crashing (#774)
• A file with both a charset declaration and BOM/SIG did not verify first the BOM/SIG charset.
• iso2022* cases misdetected due to a flaw in our multibyte chunking logic.

• Replaced the optional mypyc build with Cython extensions while retaining the pure Python fallback. The previous engine (mypyc) started to hit rough limit around the optimization of our noise/coherence detector while Cython allows us to steer the engine toward the right generated optimized sources. This change SHOULD not impact bundler (e.g. Pyinstaller) as the module are immediately discoverable (i.e. not hidden import like mypyc did). Moreover, a long wished distribution is the abi3 wheels, this will allow us to no longer rush each year when a new Python interpreter is released. We still distribute the interpreter specific wheels for faster performance.
• Applied micro-optimization on several utils.
• CharsetMatches no longer sort on each match insertion.

• Removed an old performance optimization attempt in apy.py (success_fast_tracked+payload_result_cache).

• Regression in our fallback path leading to a decode error. (#771) We've yanked 3.4.8 as a result of that bug.

• Wall import time due to cascade codec imports for our multibyte first sort of iana supported codecs (#742)
• Unnecessary json import at runtime (#753)
• Inverse capitalization not seen by noise detector (#731)

• No longer holding a global cache for our noise / coherence measurements. Relax RSS memory usage.
• Micro-optimizations in our noise / coherence measurements.
• No longer using regex search by default for our preemptive charset mark algorithm.
• Raised upperbound of setuptools to v83.
• Raised upperbound of mypy(c) to v2.1.

• Redundant UTF7 BOM marker (#730)

• Pre-built optimized version using mypy[c] v1.20.
• Relax setuptools constraint to setuptools>=68,<82.1.

• Correctly remove SIG remnant in utf-7 decoded string. (#718) (#716)

• Flattened the logic in charset_normalizer.md for higher performance. Removed eligible(..) and feed(...) in favor of feed_info(...).
• Raised upper bound for mypy[c] to 1.20, for our optimized version.
• Updated UNICODE_RANGES_COMBINED using Unicode blocks v17.

• Edge case where noise difference between two candidates can be almost insignificant. (#672)
• CLI --normalize writing to wrong path when passing multiple files in. (#702)

• Freethreaded pre-built wheels now shipped in PyPI starting with 3.14t. (#616)

• Update setuptools constraint to setuptools>=68,<=82.
• Raised upper bound of mypyc for the optional pre-built extension to v1.19.1

• Add explicit link to lib math in our optimized build. (#692)
• Logger level not restored correctly for empty byte sequences. (#701)
• TypeError when passing bytearray to from_bytes. (#703)

• Applied safe micro-optimizations in both our noise detector and language detector.
• Rewrote the query_yes_no function (inside CLI) to avoid using ambiguous licensed code.
• Added cd.py submodule into mypyc optional compilation to reduce further the performance impact.

• Bound setuptools to a specific constraint setuptools>=68,<=81.
• Raised upper bound of mypyc for the optional pre-built extension to v1.18.2

• setuptools-scm as a build dependency.

• Enforced hashes in dev-requirements.txt and created ci-requirements.txt for security purposes.
• Additional pre-built wheels for riscv64, s390x, and armv7l architectures.
• Restore multiple.intoto.jsonl in GitHub releases in addition to individual attestation file per wheel.

• mypy(c) is no longer a required dependency at build time if CHARSET_NORMALIZER_USE_MYPYC isn't set to 1. (#595) (#583)
• automatically lower confidence on small bytes samples that are not Unicode in detect output legacy function. (#391)

• Custom build backend to overcome inability to mark mypy as an optional dependency in the build phase.
• Support for Python 3.14

• sdist archive contained useless directories.
• automatically fallback on valid UTF-16 or UTF-32 even if the md says it's noisy. (#633)

• SBOM are automatically published to the relevant GitHub release to comply with regulatory changes. Each published wheel comes with its SBOM. We choose CycloneDX as the format.
• Prebuilt optimized wheel are no longer distributed by default for CPython 3.7 due to a change in cibuildwheel.

• Addressed the DeprecationWarning in our CLI regarding argparse.FileType by backporting the target class into the package. (#591)
• Improved the overall reliability of the detector with CJK Ideographs. (#605) (#587)

• Optional mypyc compilation upgraded to version 1.15 for Python >= 3.8

• Project metadata are now stored using pyproject.toml instead of setup.cfg using setuptools as the build backend.
• Enforce annotation delayed loading for a simpler and consistent types in the project.
• Optional mypyc compilation upgraded to version 1.14 for Python >= 3.8

• pre-commit configuration.
• noxfile.

• build-requirements.txt as per using pyproject.toml native build configuration.
• bin/integration.py and bin/serve.py in favor of downstream integration test (see noxfile).
• setup.cfg in favor of pyproject.toml metadata configuration.
• Unused utils.range_scan function.

• Converting content to Unicode bytes may insert utf_8 instead of preferred utf-8. (#572)
• Deprecation warning "'count' is passed as positional argument" when converting to Unicode bytes on Python 3.13+

• Argument --no-preemptive in the CLI to prevent the detector to search for hints.
• Support for Python 3.13 (#512)

• Relax the TypeError exception thrown when trying to compare a CharsetMatch with anything else than a CharsetMatch.
• Improved the general reliability of the detector based on user feedbacks. (#520) (#509) (#498) (#407) (#537)
• Declared charset in content (preemptive detection) not changed when converting to utf-8 bytes. (#381)

• Unintentional memory usage regression when using large payload that match several encoding (#376)
• Regression on some detection case showcased in the documentation (#371)

• Noise (md) probe that identify malformed arabic representation due to the presence of letters in isolated form (credit to my wife)

• Optional mypyc compilation upgraded to version 1.6.1 for Python >= 3.8
• Improved the general detection reliability based on reports from the community

• Allow to execute the CLI (e.g. normalizer) through python -m charset_normalizer.cli or python -m charset_normalizer
• Support for 9 forgotten encoding that are supported by Python but unlisted in encoding.aliases as they have no alias (#323)

• (internal) Redundant utils.is_ascii function and unused function is_private_use_only
• (internal) charset_normalizer.assets is moved inside charset_normalizer.constant

• (internal) Unicode code blocks in constants are updated using the latest v15.0.0 definition to improve detection
• Optional mypyc compilation upgraded to version 1.5.1 for Python >= 3.8

• Unable to properly sort CharsetMatch when both chaos/noise and coherence were close due to an unreachable condition in __lt__ (#350)

• Typehint for function from_path no longer enforce PathLike as its first argument
• Minor improvement over the global detection reliability

• Introduce function is_binary that relies on main capabilities, and optimized to detect binaries
• Propagate enable_fallback argument throughout from_bytes, from_path, and from_fp that allow a deeper control over the detection (default True)
• Explicit support for Python 3.12

• Edge case detection failure where a file would contain 'very-long' camel cased word (Issue #289)

• Argument should_rename_legacy for legacy function detect and disregard any new arguments without errors (PR #262)

• Support for Python 3.6 (PR #260)

• Optional speedup provided by mypy/c 1.0.1

• Multi-bytes cutter/chunk generator did not always cut correctly (PR #233)

• Speedup provided by mypy/c 0.990 on Python >= 3.7

• Extend the capability of explain=True when cp_isolation contains at most two entries (min one), will log in details of the Mess-detector results
• Support for alternative language frequency set in charset_normalizer.assets.FREQUENCIES
• Add parameter language_threshold in from_bytes, from_path and from_fp to adjust the minimum expected coherence ratio
• normalizer --version now specify if current version provide extra speedup (meaning mypyc compilation whl)

• Build with static metadata using 'build' frontend
• Make the language detection stricter
• Optional: Module md.py can be compiled using Mypyc to provide an extra speedup up to 4x faster than v2.1

• CLI with opt --normalize fail when using full path for files
• TooManyAccentuatedPlugin induce false positive on the mess detection when too few alpha character have been fed to it
• Sphinx warnings when generating the documentation

• Coherence detector no longer return 'Simple English' instead return 'English'
• Coherence detector no longer return 'Classical Chinese' instead return 'Chinese'
• Breaking: Method first() and best() from CharsetMatch
• UTF-7 will no longer appear as "detected" without a recognized SIG/mark (is unreliable/conflict with ASCII)
• Breaking: Class aliases CharsetDetector, CharsetDoctor, CharsetNormalizerMatch and CharsetNormalizerMatches
• Breaking: Top-level function normalize
• Breaking: Properties chaos_secondary_pass, coherence_non_latin and w_counter from CharsetMatch
• Support for the backport unicodedata2

• Extend the capability of explain=True when cp_isolation contains at most two entries (min one), will log in details of the Mess-detector results
• Support for alternative language frequency set in charset_normalizer.assets.FREQUENCIES
• Add parameter language_threshold in from_bytes, from_path and from_fp to adjust the minimum expected coherence ratio

• Build with static metadata using 'build' frontend
• Make the language detection stricter

• CLI with opt --normalize fail when using full path for files
• TooManyAccentuatedPlugin induce false positive on the mess detection when too few alpha character have been fed to it

• Coherence detector no longer return 'Simple English' instead return 'English'
• Coherence detector no longer return 'Classical Chinese' instead return 'Chinese'

• normalizer --version now specify if current version provide extra speedup (meaning mypyc compilation whl)

• Breaking: Method first() and best() from CharsetMatch
• UTF-7 will no longer appear as "detected" without a recognized SIG/mark (is unreliable/conflict with ASCII)

• Sphinx warnings when generating the documentation

• Optional: Module md.py can be compiled using Mypyc to provide an extra speedup up to 4x faster than v2.1

• Breaking: Class aliases CharsetDetector, CharsetDoctor, CharsetNormalizerMatch and CharsetNormalizerMatches
• Breaking: Top-level function normalize
• Breaking: Properties chaos_secondary_pass, coherence_non_latin and w_counter from CharsetMatch
• Support for the backport unicodedata2

• Function normalize scheduled for removal in 3.0

• Removed useless call to decode in fn is_unprintable (#206)

• Third-party library (i18n xgettext) crashing not recognizing utf_8 (PEP 263) with underscore from @aleksandernovikov (#204)

• Output the Unicode table version when running the CLI with --version (PR #194)

• Reuse decoded buffer for single byte character sets from @nijel (PR #175)
• Fixing some performance bottlenecks from @deedy5 (PR #183)

• Workaround potential bug in cpython with Zero Width No-Break Space located in Arabic Presentation Forms-B, Unicode 1.1 not acknowledged as space (PR #175)
• CLI default threshold aligned with the API threshold from @oleksandr-kuzmenko (PR #181)

• Support for Python 3.5 (PR #192)

• Use of backport unicodedata from unicodedata2 as Python is quickly catching up, scheduled for removal in 3.0 (PR #194)

• ASCII miss-detection on rare cases (PR #170)

• Explicit support for Python 3.11 (PR #164)

• The logging behavior have been completely reviewed, now using only TRACE and DEBUG levels (PR #163 #165)

• Fallback match entries might lead to UnicodeDecodeError for large bytes sequence (PR #154)

• Skipping the language-detection (CD) on ASCII (PR #155)

• Moderating the logging impact (since 2.0.8) for specific environments (PR #147)

• Wrong logging level applied when setting kwarg explain to True (PR #146)

• Improvement over Vietnamese detection (PR #126)
• MD improvement on trailing data and long foreign (non-pure latin) data (PR #124)
• Efficiency improvements in cd/alphabet_languages from @adbar (PR #122)
• call sum() without an intermediary list following PEP 289 recommendations from @adbar (PR #129)
• Code style as refactored by Sourcery-AI (PR #131)
• Minor adjustment on the MD around european words (PR #133)
• Remove and replace SRTs from assets / tests (PR #139)
• Initialize the library logger with a NullHandler by default from @nmaynes (PR #135)
• Setting kwarg explain to True will add provisionally (bounded to function lifespan) a specific stream handler (PR #135)

• Fix large (misleading) sequence giving UnicodeDecodeError (PR #137)
• Avoid using too insignificant chunk (PR #137)

• Add and expose function set_logging_handler to configure a specific StreamHandler from @nmaynes (PR #135)
• Add CHANGELOG.md entries, format is based on Keep a Changelog (PR #141)

• Add support for Kazakh (Cyrillic) language detection (PR #109)

• Further, improve inferring the language from a given single-byte code page (PR #112)
• Vainly trying to leverage PEP263 when PEP3120 is not supported (PR #116)
• Refactoring for potential performance improvements in loops from @adbar (PR #113)
• Various detection improvement (MD+CD) (PR #117)

• Remove redundant logging entry about detected language(s) (PR #115)

• Fix a minor inconsistency between Python 3.5 and other versions regarding language detection (PR #117 #102)

• Unforeseen regression with the loss of the backward-compatibility with some older minor of Python 3.5.x (PR #100)
• Fix CLI crash when using --minimal output in certain cases (PR #103)

• Minor improvement to the detection efficiency (less than 1%) (PR #106 #101)

• The project now comply with: flake8, mypy, isort and black to ensure a better overall quality (PR #81)
• The BC-support with v1.x was improved, the old staticmethods are restored (PR #82)
• The Unicode detection is slightly improved (PR #93)
• Add syntax sugar __bool__ for results CharsetMatches list-container (PR #91)

• The project no longer raise warning on tiny content given for detection, will be simply logged as warning instead (PR #92)

• In some rare case, the chunks extractor could cut in the middle of a multi-byte character and could mislead the mess detection (PR #95)
• Some rare 'space' characters could trip up the UnprintablePlugin/Mess detection (PR #96)
• The MANIFEST.in was not exhaustive (PR #78)

• The CLI no longer raise an unexpected exception when no encoding has been found (PR #70)
• Fix accessing the 'alphabets' property when the payload contains surrogate characters (PR #68)
• The logger could mislead (explain=True) on detected languages and the impact of one MBCS match (PR #72)
• Submatch factoring could be wrong in rare edge cases (PR #72)
• Multiple files given to the CLI were ignored when publishing results to STDOUT. (After the first path) (PR #72)
• Fix line endings from CRLF to LF for certain project files (PR #67)

• Adjust the MD to lower the sensitivity, thus improving the global detection reliability (PR #69 #76)
• Allow fallback on specified encoding if any (PR #71)

• Part of the detection mechanism has been improved to be less sensitive, resulting in more accurate detection results. Especially ASCII. (PR #63)
• According to the community wishes, the detection will fall back on ASCII or UTF-8 in a last-resort case. (PR #64)

• Empty/Too small JSON payload miss-detection fixed. Report from @tseaver (PR #59)

• Don't inject unicodedata2 into sys.modules from @akx (PR #57)

• Make it work where there isn't a filesystem available, dropping assets frequencies.json. Report from @sethmlarson. (PR #55)
• Using explain=False permanently disable the verbose output in the current runtime (PR #47)
• One log entry (language target preemptive) was not show in logs when using explain=True (PR #47)
• Fix undesired exception (ValueError) on getitem of instance CharsetMatches (PR #52)

• Public function normalize default args values were not aligned with from_bytes (PR #53)

• You may now use charset aliases in cp_isolation and cp_exclusion arguments (PR #47)

• 4x to 5 times faster than the previous 1.4.0 release. At least 2x faster than Chardet.
• Accent has been made on UTF-8 detection, should perform rather instantaneous.
• The backward compatibility with Chardet has been greatly improved. The legacy detect function returns an identical charset name whenever possible.
• The detection mechanism has been slightly improved, now Turkish content is detected correctly (most of the time)
• The program has been rewritten to ease the readability and maintainability. (+Using static typing)+
• utf_7 detection has been reinstated.

• This package no longer require anything when used with Python 3.5 (Dropped cached_property)
• Removed support for these languages: Catalan, Esperanto, Kazakh, Baque, Volapük, Azeri, Galician, Nynorsk, Macedonian, and Serbocroatian.
• The exception hook on UnicodeDecodeError has been removed.

• Methods coherence_non_latin, w_counter, chaos_secondary_pass of the class CharsetMatch are now deprecated and scheduled for removal in v3.0

• The CLI output used the relative path of the file(s). Should be absolute.

• Logger configuration/usage no longer conflict with others (PR #44)

• Using standard logging instead of using the package loguru.
• Dropping nose test framework in favor of the maintained pytest.
• Choose to not use dragonmapper package to help with gibberish Chinese/CJK text.
• Require cached_property only for Python 3.5 due to constraint. Dropping for every other interpreter version.
• Stop support for UTF-7 that does not contain a SIG.
• Dropping PrettyTable, replaced with pure JSON output in CLI.

• BOM marker in a CharsetNormalizerMatch instance could be False in rare cases even if obviously present. Due to the sub-match factoring process.
• Not searching properly for the BOM when trying utf32/16 parent codec.

• Improving the package final size by compressing frequencies.json.
• Huge improvement over the larges payload.

• CLI now produces JSON consumable output.
• Return ASCII if given sequences fit. Given reasonable confidence.

• In some very rare cases, you may end up getting encode/decode errors due to a bad bytes payload (PR #40)

• Empty given payload for detection may cause an exception if trying to access the alphabets property. (PR #39)

• The legacy detect function should return UTF-8-SIG if sig is present in the payload. (PR #38)

• Amend the previous release to allow prettytable 2.0 (PR #35)

• Fix error while using the package with a python pre-release interpreter (PR #33)

• Dependencies refactoring, constraints revised.

• Add python 3.9 and 3.10 to the supported interpreters