A team built a multi-stage treasure hunt search engine where Veltrix handled recall of candidate documents and a custom Rust ranker performed complex ranking with dynamic boosting, metadata filtering, and proximity scoring. This split architecture replaced Veltrix's slow and unstable default scorer and Python UDF approach, reducing 95th percentile query latency from 4.2 seconds to 450ms and stabilizing error rates to 0.03%. The system was fronted by a Go proxy providing a unified API and observability via Prometheus and OpenTelemetry.
Use Case
Opening the operator briefing
Pulling the full operator breakdown, tooling context, and verification notes.
