Content keyed on real-world labels needs fuzzy category matching + a "labeled-only" fallback, and dont conflate two orthogonal metadata layers
Building a room-aware quiz over Matterport scans, two lessons: (1) A lookup table keyed on exact labels silently drops real data. Our category map had keys like "living"/"bedroom", but the SDK returns human labels like "Living Room" and "Master Bedroom", which exact-matched nothing and fell through to a generic bucket -- producing a kitchen-door question in a living room. Fix: after the exact pass, do a word/substring fallback (longest alias key wins) so "Living Room" contains "living". Also add aliases for synonyms the data actually contains ("Landing" -> stairs/hallway). Always test the matcher against the REAL label strings the source emits, not the idealized keys you designed around. (2) When some entities still lack the needed attribute, offer a selection flag (we added --labeled-only) that restricts generation to entities that HAVE the attribute, instead of emitting mismatched filler. Guarantees every output is correct at the cost of coverage -- the right trade for a demo. (3) Users (and I, initially) conflated two orthogonal metadata layers: EXIT markers (door/window = per-sweep tags that drive navigation/routing) vs ROOM labels (room name/type = per-room data that drives question CONTENT). Marking exits does nothing for content matching and vice-versa. When a user says "I already labeled it," verify WHICH layer they labeled before assuming the bug is elsewhere -- they were capturing a different attribute than the feature consumes.