Retail Search

ARCH-0.5 · cross-domain core

 Live · Amazon ESCI

Phase 3 - Retail Search Explain

How products rank. ESCI products are structured (title, brand, color, bullets, description). BM25 scores each field with per-field boosts (title x3, brand x2); the BGE hybrid adds semantic matching via a runtime query embedding. Run a query to see which fields made each product rank.

Try searchAll phases

Explain a product ranking (field-level)

Run explain to see the per-field score breakdown for the top products.

Retail ranking flow reference

  1. 1Structured fieldstitle, brand, color, bullets, description indexed separately
  2. 2Per-field BM25best_fields with boosts: title x3, brand x2, color/bullets/description x1
  3. 3Semantic (hybrid only)Workers AI BGE query embedding -> int8 kNN over 1.2M product vectors
  4. 4Blendmin-max linear blend of lexical + semantic (hybrid); graded ESCI labels E/S/C/I score the ranking
  5. 5Negation query understanding (hybrid)parse "without X / X-free" constraints; demote candidates that positively contain X, protecting products that advertise the absence - the one failure class neither lexical nor dense retrieval solved
Explain JSONEvaluation JSON