0.2.0

What's Changed

Happy new year! This release contains a significant overhaul of many elements of guidance, including:

• Visualization Overhaul: Guidance's in-line visualizations have been rewritten from scratch. TL;DR: We added a lot of code to support better model metrics and visualization. Check it out:

• Significantly Faster, Rust-based Grammar Engine: Guidance's core grammar processing has been fully migrated to the llguidance Rust library. Our grammar processing is now significantly faster -- state of the art across frameworks -- and fixes some key, subtle bugs in the earlier processing engine. All constrained decoding code will now run much faster.

• Increased JSON schema coverage: We've spent a lot of time focusing on JSON schema support in guidance, and now have leading coverage of JSON schema constraint types via guidance's json function.

• And lots more general quality of life improvements! See below for full list of PRs and changelogs. Thank you to all the existing and new contributors that made this possible!

---

• [WIP] Replace EarleyParser with lexeme-based rust implementation by @hudson-ai in #951
• Update README.md with new grammar-enabled endpoint details (Azure AI) by @Harsha-Nori in #998
• [Test] Enable Azure Guidance endpoints in test matrix by @riedgar-ms in #999
• [Test] Move Azure Guidance in CI by @riedgar-ms in #1000
• [Bug] del f._self_call_placeholder_ even when calling f raises exception by @hudson-ai in #994
• [mypy] Improve types in _grammar.py by @riedgar-ms in #1002
• [Build] CI & Notebook Workaround by @riedgar-ms in #1005
• [Feature] Allow oneOf in JSON schemas (with limited support) by @hudson-ai in #982
• Use json_string kwarg from llguidance 0.1.7to properly escape json string patterns by @hudson-ai in #1008
• Remove Placeholder and replace_grammar_node in favor of pointer-like container object by @hudson-ai in #1007
• [Feature] Support required properties in JSON schemas by @hudson-ai in #1009
• [Housekeeping] Clean up Enums to make JSON features more transparent by @hudson-ai in #1011
• [Doc] Fix typos in README and tutorials notebooks by @gfesatidis in #1012
• fix test_ll for new llguidance with working advance_parser() by @mmoskal in #1013
• [Bug] Ensure grammar.match raises appropriately on incomplete input by @hudson-ai in #1016
• Support boolean JSON schemas by @hudson-ai in #1015
• @overload the guidance decorator in order to improve static type-checking by @hudson-ai in #1014
• JSON string 'format' kwarg by @hudson-ai in #984
• Strip anchors from JSON regexes by @hudson-ai in #1027
• JSON reserved keywords by @hudson-ai in #1032
• "required" JSON keys are validated against "additionalProperties" if they are missing from "properties" by @hudson-ai in #1029
• [Bug] Make _self_call_placeholder_ thread-safe by @hudson-ai in #1034
• add test for "nullable bug" by @mmoskal in #1025
• Make the tests pass with larger tokenizers by @mmoskal in #1037
• [Feature] Support for JSON minimum , maximum, exclusiveMinimum, and exclusiveMaximum by @hudson-ai in #1023
• [Feature] Allow @guidance to decorate methods by @hudson-ai in #1035
• [Bug] [Feature] Fix issue #1039 by exposing And, Not regex operators from llguidance by @hudson-ai in #1043
• [WIP] Try to get PR Gate working again... by @hudson-ai in #1048
• [Feature] Enable user-passed separators keyword to guidance.json by @hudson-ai in #1044
• fix typos in README by @Saibo-creator in #1056
• [Feature] Expand JSON $ref resolution by @hudson-ai in #1061
• [JSON] Raise exceptions when "sibling" keywords are unhandled by @hudson-ai in #1063
• [Feature] Allow json-loads-able strings to be passed as schema by @hudson-ai in #1028
• [Hotfix] Echo forced to false on terminal detection. by @nopdive in #1067
• [Keeping House] Split JSON tests into multiple files by @hudson-ai in #1072
• [Feature] [JSON] Broader support of allOf JSON keyword by @hudson-ai in #1070
• [BUG] Reset role blocks when model.reset is called by @hudson-ai in #1073
• [Optimization] Advance parser concurrently with model forward pass by @hudson-ai in #1065
• Update README.md by @culmat in #1045
• [Feature] Add qwen2.5 chat templates by @xruifan in #1066
• Expose fastforward and backtrack llguidance flags by @Harsha-Nori in #1075
• [Bug] Fix for environment detection in barebones environments. by @nopdive in #1076
• Update setup.py to limit llguidance version by @hudson-ai in #1080
• update to llguidance 0.4.1 by @mmoskal in #1081
• fix current failures in pytest when run locally by @paulbkoch in #1083
• Force Transformers engine to return float logits by @JC1DA in #1086
• Bump runners from macos-12 to macos-14 by @hudson-ai in #1088
• update llguidance to 0.5.0 by @mmoskal in #1087
• Use llguidance JSON grammar by @hudson-ai in #1085
• Upgrade llguidance to 0.5.1rc0 by @hudson-ai in #1090
• Interactivity Overhaul (User Interface & Model Instrumentation & Network Comms) by @nopdive in #1054

New Contributors

• @gfesatidis made their first contribution in #1012
• @Saibo-creator made their first contribution in #1056
• @culmat made their first contribution in #1045
• @xruifan made their first contribution in #1066
• @JC1DA made their first contribution in #1086

Full Changelog: 0.1.16...0.2.0