Match & dedupe products / records — the best accuracy for the cost, benchmarked honestly.
Matching two catalogs, deduping records, linking product offers? Your choices today are: a free fuzzy
library (fast, but low accuracy), or an LLM call per pair (accurate, but slow and expensive at scale). We
sit on the frontier between them: near-LLM accuracy at fuzzy-library speed and cost, as a
single MCP tool an agent calls — match_records(a, b) → score.
| method | F1 tuned to domain (Abt-Buy, val.) | F1 zero-shot (held-out Walmart test) ✓ | $/1M pairs | latency |
|---|---|---|---|---|
| rapidfuzz (free) | 0.485 | 0.414 | ~$0 | ~0.01 ms |
| Kywio Match | 0.860 | 0.624 | $0.0012 | <1 ms |
| cheap LLM (gpt-4o-mini) | 0.903 | 0.774 | $15–20 | ~600 ms |
Honest read: we beat free fuzzy libraries on hard textual/product matching (on short, distinctive names like beer, simple fuzzy can match us), and we cost ~15,000× less and run ~750× faster than an LLM. Tuned to your domain we're near LLM accuracy; out-of-the-box there's a real accuracy gap to an LLM — we show it rather than hide it. ✓ = a verified one-shot held-out test (never-seen Walmart-Amazon, thresholds frozen on validation); the tuned column is validation. Method + numbers: https://github.com/vanekyj/kywio