v0.5.0

• Backends
  • ClaudeAPI: Added new backend for Anthropic Claude models, including
    support for tool use and thinking control.
  • OpenAI API: Migrated to the latest Responses API, adding support for
    reasoning, tools, .embed, and .count_tokens.
  • GoogleGenAIAPI:
    • Updated default Gemini API model to gemini-3.1-flash-lite.
    • Enabled server-side prefix caching.
    • Added CachedChunkList for robust prefix cache handling.
    • Removed token healing functionality.
    • Added normalize kwarg to .embed.
    • Improved token usage reporting in tracers.
    • Added support for Gemini Function Calling (FC2.0) metadata.
    • Bug fixes and model updates.
  • SentencePieceTokenizer: Added a new tokenizer backend using the
    standard sentencepiece library.
• Core
  • Execution & Parallelism: Fixed event loop leak in
    asyncio_run_wrapper.
  • Tracing:
    • Improved exception handling and reporting within the tracing system.
    • Reduced tracing overhead by skipping large inputs/outputs in traces
      for several backends and modules.
• Standard library
  • Retrieval & QA:
    • Architectural split for constrained retrieval, introducing
      ConstrainedEmbeddingBasedIndex.
    • Removed use of lambdas in EmbeddingBasedIndex for better pickle
      support.
    • Refined chunking logic in ChunkByMaxTokens.
    • Added QAPromptBuilder for flexible QA prompt generation.
  • code_execution: Fixed test for Python 3.14.
• Evaluation
  • ot.evaluate: Optimized to write large results and results_debug
    JSON files incrementally to avoid OOM errors.
• Colabs & Docs
  • RAG Tutorial: Significantly revamped with improved explanations and
    diagrams.
  • Tutorial: Updated with ClaudeAPI examples.