Distillation

Date: Sep 16, 2026

• 🔴 (23:44) User stated embedding-worker recovery uses current container free memory so that climbing the token cap back up never lets the aggregate memory of N workers exceed the host’s available capacity.
• 🔴 (23:44) User stated the cap that caused an embedding OOM is remembered, and upward re-probing never climbs back to that failed cap.
• 🔴 (23:44) User stated upward re-probing is anchored to free memory measured at OOM time, permitting a climb only when memory headroom improves.
• 🟡 (23:44) The embedding worker clears stale worker errors before deliberate respawn so the replacement worker is not immediately failed by obsolete state.
• 🟡 (23:44) When no worker can settle pending embedding requests, the requests are explicitly rejected rather than left hanging.
• 🟡 (23:44) Missing optional local-embedding support is latched as a degraded state that does not self-heal; recall falls back to FTS-only search, with guidance to reinstall without --omit=optional or configure a remote provider in .lore.json using VOYAGE_API_KEY or OPENAI_API_KEY.
• 🟡 (23:44) Local embedding initialization retries transient failures but eventually latches failure for the process lifetime rather than retrying indefinitely.
• 🟡 (23:44) terminate() emits an asynchronous exit(1) from the terminated worker, so deliberate worker shutdown and respawn must account for that event.
• 🟡 (23:44) Synchronous worker-respawn failure rejects outstanding requests because no later worker event can settle them.
• 🟡 (23:44) embed(...) pre-truncates texts that exceed the safe ONNX inference limit before dispatching them to the worker.