• Minor improvements required for vLLM V1 Engine support

• Fixed support for newer versions of transformer library

• Added support of multiple types for a field in json schema (Contributed by Guillaume Calmettes)
• Allowing configurating the alphabet via environment variable (Addresses #151)
• Fixed CI build issues

• Added new sample: Extracting structured data from vision models

• Fixed support for tokenizers with larger vocabularies than the models. Enables visual models such as Llama3.2
• JsonSchemaParser: Fixed additionalProperies and oneOf edge cases

v0.10.7

• [135] Updated Haystack V2 integration with latest API

v0.10.6

• Pickling improvements for easier multiprocessing support

v0.10.5

• SequenceParser performance optimization
• JsonSchemaParser: number parsing supports exponents
• Supporting tokenizers with multiple eos token ids

v0.10.4

• Added default max Json array length to help LLMs avoid infinite loops. See README for details.
• Updated EXLlamaV2 example to updated API

v0.10.3

• [#113] TRTLLM Support: Fixing type incompatibility in certain cases / library versions

• [#100] JsonSchemaParser: Added allOf support
• [#99] JsonSchemaParser: Fixed edge case that would allow leading comma in JSON Array
• [#102] JsonSchemaParser: Fixed Array of Enums not producing multiple values

• Added ability to config parsing heuristics via environment variables. See Configuration Options in README
• JsonSchemaParser: Added support for strict json field order and max consecutive whitespace configuration overrides.

• [#95] Added anyOf support to JsonSchemaParser, making function calls possible.

• Updated README with vLLM OpenAI Server Inference integration

• [#80] JSONSchemaParser List would allow opening comma before first element if there was a whitespace before it

v0.9.7

• [#93] Improved JSONSchemaParser performance, unit tests run twice as fast! Joint effort with Ari Weinstein. Thanks!

v0.9.6

• [#88] ExllamaV2 optimizations
• Bugfix in ExllamaV2 sample notebook that generated garbage data after the response.

• [#87] Allow regex to exit to forcestopparser when receiving a pad/eos token after having been in a final state. Thanks Josh C!

• [#27] Improving vLLM class support (AsyncLLMEngine etc)

• #83 - Supporting added model tokens
• Improved support for out-of-token-vocabulary characters

• #80 - Fixed bug where comma could start json list
• #34 - Fixed llama-cpp-python low max_tokens default in sample

• Fixed build errors in certain edge cases

• #68 Added NVIDIA TensorRT-LLM Support, NVIDIA's contribution by Ahmet Erdem. Thanks!
• Much faster TokenizerData initialization, new JSON freetext token caching algorithm.
• More robust error reporting.

• #67 Updating vLLM integration to support v0.3.0
• #63 JSONSchemaParser: Empty list cannot be closed after a newline

Several JsonSchemaParser improvements:

• #32 Added limited support for regex-constrained string in JSON using the pattern field. See test_phone_number_in_string().
• #54 Fixed regression bug caused by limited-length JSON string caching.
• #53 Fixed problems with arrays of union types.

• Performannce improvement: Limited-length JSON strings will also enjoy caching. Thanks Jarno Elonen for the contribution!

• Performance improvement: Introduced TokenEnforcerTokenizerData that allows reusing the tokenizer preprocessing data between different TokenEnforcer instances. The sample notebooks have been updated to take advantage of this option.
• Performance improvement: Long sequences will see up to 5x TokenEnforcer runtime footprint reduction.

• Bug fixes

• vLLM performance improvements
• Sample notebooks don't require huggingface account anymore

• Added ExLlamaV2 integration

• JSON Schema: Added support for union types. In pydantic, both key: int | str and key: Union[int, str] formats are supported
• JSON Schema: Added support for schemaless JSON mode. JsonSchemaParser(None) will now create a parser that accepts any valid JSON.

• Added official vLLM integration that doesn't require monkey patching.

• JSON Schema : Supports string min/max length limitation

• Community PR: Fixed SequenceParser bug

• Added haystack integration

• Fixed llama.cpp integration to be able to generate unicode characters in json freetext fields
• Fixed unescaped newlines being allowed in json freetext fields

• RegexParser and JsonSchemaParser can now output all of the characters that exist in the tokenzier
• Added "Known issues and limitations" section to the README
• Fixed a bug in JsonSchemaParser where sometimes illegal commas were allowed

• JSON Schema : Supports empty arrays and escape characters in strings
• Regex : Performance improvement in some cases
• Added UnionParser and SequenceParser to allow combining parsers

• Made it easier to report bugs in the library

• Introduced FormatEnforcerAnalyzer to allow all inference engines to be analyzed in a unified way. (Was previously only available for transformers)
• Added support for the analyser in llama.cpp, updated example notebook
• JsonSchemaParser now take list min/max items into consideration

• Improved JsonSchemaParser whitespace support
• Improved RegexParser performance, especially in regular expressions with .+ and .* sections.

• Modified example in main README to be able to run end to end in Google Colab

• Added integration with the LlamaIndex library (huggingface and llama.cpp backends) via sample notebook.

• Introduced lmformatenforcer.integrations module which will have the integrations with inference engines.
• Added llama-cpp-python integration to the library in lmformatenforcer.integrations.llamacpp
• Breaking API Change: Moved 'build_transformers_prefix_allowed_tokens_fn', 'generate_enforced' from lmformatenforcer to lmformatenforcer.integrations.transformers.