Memory
DocuClear ingestion service decisions: adapt clearspace-curator into new repo docuclear-ingestion (adapt, never rebuild — the findContent fallback and polling tuning are paid-for debugging). Return-content pattern: service never writes customer stores. Shared container pool for xberg/crawlberg — NO per-customer containers, NO VPS/tunnel phase (containers bill only while awake, sleepAfter). Per-slug CuratorAgent DOs via idFromName(slug). Auth: platform secret + X-Sphere-Slug mandatory, no bypass headers. INGESTION_CONTRACT.md identical in both repos, immutable. Sphere bundle carries a slim IngestionAgent DO (orchestration + DO-alarm scheduling, never cron). Embeddings happen IN SPHERE, model choice is plan-driven, upstream has a rebuild-vectors utility — consequence: Vectorize index dims are fixed at creation, so model change = new index + rebind path needed in platform update workflow.
Tags: kind:semantic, status:canonical — Source: mcp — 2026-08-20 11:20:12 UTC