Product Matching

Matching is a cascade. Cheap deterministic checks run first; LLM steps are reserved for borderline cases.

flowchart TD A[Candidate URL] --> B[Exact GTIN] B -->|miss| C[MPN plus brand] C -->|miss| D[Normalized name] D -->|borderline| E[Embedding semantic match] E -->|fashion or luxury| F[Visual matcher] F -->|60 to 84| G[LLM judge] G --> H{Score} H -->|85 plus| I[Auto confirmed] H -->|60 to 84| J[Human review] H -->|below 60| K[Rejected]

The acceptance bands are:

Score Result
85..100 Confirmed
60..84 Match proposal
0..59 Rejected
Cost control

The package caches embeddings and avoids re-running expensive AI steps unless candidate evidence changes materially.

previous Catalog and Targets next Marketplace Adapters