Version 2.0 brings foundational improvements to answer generation, reliability, and user control. OPENAI RESPONSES API & GPT-5 • Migrated from Chat Completions to structured Responses API • Native GPT-5 reasoning model with dynamic effort controls • Existing models (GPT-4o, GPT-4o mini) continue with temperature/top-p • Settings UI conditionally renders correct parameters • Structured "input" field with system/user content arrays SERVER-SENT EVENTS STREAMING • Answers stream token-by-token instead of blocking • Typing indicator appears immediately while waiting • Stop button cancels mid-generation with clear feedback • Watchdog timer detects stalls and triggers fallback completion CONVERSATION MEMORY • Bounded client mode includes last ~8 messages for context • Server-managed threads use OpenAI conversation IDs • Toggle between modes in Settings for privacy preferences • Follow-ups understand prior exchanges without restating RESILIENCE & CIRCUIT BREAKER • Pinecone health checks run before queries with short timeout • Circuit breaker prevents calls to unhealthy hosts • Per-search trace IDs tie logs to specific queries • Error banners auto-dismiss with user-friendly summaries PRIVACY & RESET FLOWS • Settings > Data & Privacy exposes credential reset • Clears Keychain, conversation history, bookmark consent • Onboarding messaging clarifies missing keys • Preflight script blocks accidental credential leaks CITATION STABILITY & POLISH • Citations enumerate by offset—no duplicate ID warnings • Logs tab filters with color-coded levels • Index host transitions emit structured logs • Improved keyboard handling and haptic feedback Behind the scenes: upgraded to OpenAI Responses API, added conditional model parameterization, implemented SSE parsing with cancellation, and wired bounded conversation memory. All changes maintain backward compatibility with existing documents and Pinecone indexes.