v3.3.0

This release includes significant improvements to the subtitle generation prompt, optimizations for parallel web searches, and refinement of the data models for better AI interaction.

New Features & Improvements:

• Enhanced Lyrics Research:
  • Updated LYRICS_PROMPT_VERSION to 9 and SUBTITLES_PROMPT_VERSION to 18.
  • Implemented a stricter parallel search strategy (limited to 3 turns) to reduce costs and latency.
  • Added critical constraints to prohibit manual transcription/translation and ensure official lyrics are retrieved via search.
• Data Model Refinement:
  • Optimized JSON schemas by hiding the thoughts field from the LLM using SkipJsonSchema in src/ai_sub/data_models.py.
  • Improved field descriptions for timestamps and scene content to enforce strict 'MM:SS.mmm' formatting and prevent hallucinated content.
• Pipeline Stability:
  • Simplified the SubtitleAiResponse schema by removing the global_analysis field, streamlining the generation process.
  • Refined log messaging for prompt version mismatches to improve debugging.

What's Changed

• Remove output format from prompt by @FlippFuzz in #15

Full Changelog: v3.2.0...v3.3.0