Memory
Model placement decision for docuclear-ingestion (Aug 2026): three-way split. (1) EMBEDDINGS: sphere-side via Workers AI binding (already decided — model per plan tier) — xberg's embedding models are NOT needed, never download them. (2) OCR/core extraction: stay IN the xberg container image (Tesseract data + core models baked at build) because enableInternet=false forbids runtime downloads AND CF container disk is fully ephemeral (fresh disk from image on every wake) — bake-in is the only working option. Container OCR ~$0.0001/page vs vision-model ~$0.002-0.01/page. (3) PREMIUM structured/vision extraction: Workers AI vision/LLM models called at the Worker layer (service DO with AI binding, or sphere) as pro/enterprise path — containers can't reach bindings. Crawlberg needs no models. REBUILD.md must document: image contains Tesseract + core models, no embedding models, verify with cache_manifest checksums at build.
Tags: kind:episodic, status:canonical — Source: mcp — 2026-08-20 11:56:44 UTC