Vision LLMs are great. They're also expensive at scale. So our identification flow short-circuits to cheap deterministic matchers whenever it can.
Step 1 — COA QR. Cannabis Certificates of Analysis are required by most state regulators and increasingly include a QR code linking to the lab report. Free, instant, deterministic. We hit this first; ~60% of dispensary-shelf cannabis products have a usable COA QR.
Step 2 — Terpene-sticker OCR. Some brands print a terpene panel directly on the package. Pure OCR (no LLM) extracts the cannabinoid + terpene profile.
Step 3 — Label OCR + brand match. Tesseract pulls the product name; we fuzzy-match against the local catalog scoped by tenant + dispensary location. Sub-100ms.
Step 4 — Vision LLM. Only if all three failed. We pass the photo + the dispensary's catalog as a compact prompt-cached block to Claude. Claude returns top-3 candidates with confidence; user confirms.
Cache. Perceptual hash on the photo. Re-identifying the same package twice is free.
Average cost per identify: under $0.01. Without the waterfall it would be $0.10+.