tidyllm 0.3.4

tidyllm Version 0.3.4 Release Notes

This release marks a major internal refactor accompanied by a suite of subtle yet impactful improvements. While many changes occur under the hood, they collectively deliver a more robust, flexible, and maintainable framework.

Key Improvements

• Robust Streaming:

  • New Streaming Backend: Streaming is now handled via httr2::req_perform_connection() (httr2 ≥ 1.1.1), resulting in a more stable and reliable experience.
  • Metadata for Streaming: Streaming requests now also support metadata extraction, logprobs, and other features, making them even more informative.

• Optimized Internal Processing:

  • S7 Methods Integration: Improved handling of streams and chat parsing using more proper S7 methods instead of clunky old function generation.
  • OpenAI Request Construction: Both OpenAI and Azure OpenAI (along with their batch functions) now use a common request construction function to reduce code duplication and simplify maintenance.

• Schema support:

• New field_object() function to allow for nested schemata

• Expanded API Features:

  • JSON Schema Support:
    • mistral() now accepts the .json_schema argument.
    • claude() incorporates .json_schema via a JSON-extractor tool, in line with Anthropic's guidelines.
  • Batch API for groq(): A new batch processing interface has been implemented for groq().

Bug Fixes

• claude() Batch Requests: Fixed an issue where system prompts were not transmitted correctly in batch mode.
• gemini() Prompt Handling: Resolved a bug causing system prompts to be omitted from API calls in older versions.