Intelligence Pipeline Deep Dive
What looks like a simple product list is actually the output of a multi-stage AI pipeline — personalized ranking, dietary enforcement, intent classification, and real-time reranking — all executing in under 2 seconds.
Click "Run Query" to watch a real query flow through every AI stage. Each stage transforms, filters, or re-ranks the results.
Raw natural language from the shopper
Gemini classifies the query into one of 8 intent types and extracts structured parameters.
Extract food concepts, detect false-friends, simplify compound queries, identify category hints.
Parallel load of propensity scores, virtual pantry, household graph, and purchase history.
Semantic search across 4.5M+ enriched products. Over-fetches to allow for dietary filtering downstream.
Stemming, false-friend detection, and content-word overlap scoring. Removes irrelevant matches.
Enforces dietary restrictions with AND logic. Safety-critical for allergies — never shows non-compliant products.
Reranks products based on individual shopper's 8-propensity profile and health consciousness score.
Boosts familiar brands and fresh produce. Uses purchase history and shelf-life data.
Prevents showing products the shopper already saw, selected, or dismissed in this session.
Injects relevant sponsored products at configured positions. Only shows brand-safe, dietary-compliant items.
Final ranked, filtered, personalized product set streamed to the shopper via SSE.