v0.3.4

Changed

• Voice messages now use native Gemini audio understanding instead of Yandex SpeechKit transcription
  • Audio is sent directly to the LLM as base64-encoded OGG Opus
  • Supports any language Gemini understands (not limited to Russian)
  • No external speech recognition service required
  • Yandex SpeechKit integration is now optional/unused
  • Includes explicit instruction for LLM to respond in configured language and avoid reasoning leakage

Added

• Added laplaced_build_info{version, go_version} metric for version tracking in Grafana

Fixed

• Fixed missing user_id label on LLM and Embedding metrics (missed in v0.3.3):
  • LLM: laplaced_llm_request_duration_seconds, laplaced_llm_requests_total
  • Embedding: laplaced_embedding_request_duration_seconds, laplaced_embedding_requests_total, laplaced_embedding_tokens_total, laplaced_embedding_cost_usd_total
  • Bot: laplaced_bot_context_tokens_by_source
• Fixed laplaced_memory_topics_total metric always showing 0
• Fixed histogram buckets for laplaced_bot_message_processing_duration_seconds: extended from 30s to 120s max (LLM responses can take up to a minute)