DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Mixing numeric attributes into text search for better first-stage relevance

Quality: 8/10 Relevance: 9/10

Summary

This article details how turbopuffer extends first-stage search by ranking with numeric and date attributes, using sigmoid-based scaling and the MAXSCORE engine to blend non-text signals with BM25. It covers multi-stage search architecture, latency-recall tradeoffs, and concrete query examples for ranking by attributes, along with guidance on scaling to very large corpora.

🚀 Service construit par Johan Denoyer