01
Inversion
"Which dishes contain walnuts" is written in no document — it is an inversion of 60 separate ingredient lists. The graph answers it; top-k retrieval over those lists cannot.
Spinning the strand
explore
Spinning the strand
explore
GraphRAG · 60 dishes · 121 ingredients
Sixty dishes had no single USDA record, so each was composed from weighted ingredient records. Those weights are edges. What began as a workaround for missing data is the structure the retrieval system reasons over — and it is literally a web.
drag to orbit · scroll to zoom · click a node
01
"Which dishes contain walnuts" is written in no document — it is an inversion of 60 separate ingredient lists. The graph answers it; top-k retrieval over those lists cannot.
02
58 ingredients appear in more than one dish. Those are the strands that connect the web, and the reason similarity can be computed from recipes rather than from how two dishes happen to be described.
03
Every edge carries the USDA FoodData Central record it came from, so an answer built on this graph points at a source rather than at a plausible sentence.