v0.1.1 — Robust Generation & Reliability

What's Changed

Reliability & Retry

• Per-item retry with exponential backoff — every API call (text, audio, image) now retries up to 10 times on transient errors (timeouts, 5xx, rate limits) with exponential backoff (2s → 4s → 8s → ... up to 120s max)
• Fatal errors fail fast — auth (401) and insufficient credits (402) errors propagate immediately without wasting retries
• Increased default timeout from 30s to 90s for slow models and large batches

Audio Fixes

• Streaming read timeout — added per-chunk read timeout for audio streaming to prevent hangs when the server stops sending data
• Stream error handling — timeouts and connection drops during audio streaming are now caught and retried
• Audio validation — generated audio is validated for minimum size (~50ms) to detect corrupted/empty responses

Prompt Quality

• System/user message separation — LLM instructions are now sent as system role, user data (word, question) as user role. Prevents the model from confusing example words in the prompt with the actual input word

Defaults

• Generate photo off by default in Language mode — reduces cost and generation time; can be enabled per-session via checkbox

Installation

Download AnkiForge_v0.1.1.ankiaddon and double-click to install, or use Tools → Add-ons → Install from file in Anki.