The embedding backfill (TriggerEmbeddings + the scheduled runEmbeddings)
ran under a hardcoded 30-minute context. That is fine for a fast hosted
embedding API, but local/self-hosted embedders (e.g. Ollama on CPU) are
far slower — on a large catalog they embed only a few thousand items
before the context deadline aborts the run with "context deadline
exceeded". The job is idempotent and resumable, so progress is not lost,
but it never finishes without repeatedly re-triggering it.
Make the per-run timeout configurable via a new
`recommendations.embeddings_job_timeout` setting (default 24h), threaded
through RecommendationsConfig -> NewWorker and applied to both the manual
trigger and the cron-scheduled run. A non-positive value falls back to
24h. Default behavior is unchanged for hosted users (a full backfill
comfortably fits in 24h); local LLM users can now complete a one-shot
backfill instead of stalling.
AI-use disclosure: implemented with assistance from Claude Code.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
