Decision. Catalog enrichment is a strict cascade, and each phase may only fill the NULLs the previous one left: inherit from the producer’s portfolio → match against style families → canonical defaults (all free, deterministic) → local Ollama (mistral-small:24b) extracts from shop text, grounded by a token-overlap check → only what’s still empty reaches paid Sonnet → Python rules clean up last.
Because. The expensive model should never answer a question a lookup table or a free model already answered. The cascade turned “can we afford AI enrichment?” into a line item: ~$0.012 per wine, $60 for 5k wines, every call logged to a cost-ledger table.
Instead of. One big smart model for everything — simpler to build, impossible to budget, and silently re-answering solved questions at premium rates.