Migrating a Job Search Relevance Function

TL;DR

This paper describes the migration of a homebrewed C++ search engine to OpenSearch, aimed at preserving and improving search performance with minimal impact on business metrics, and fine-tuned a new retrieval algorithm on OpenSearch.

Abstract

In this paper, we describe the migration of a homebrewed C++ search engine to OpenSearch, aimed at preserving and improving search performance with minimal impact on business metrics. To facilitate the migration, we froze our job corpus and executed queries in low inventory locations to capture a representative mixture of high- and low-quality search results. These query-job pairs were labeled by crowd-sourced annotators using a custom rubric designed to reflect relevance and user satisfaction. Leveraging Bayesian optimization, we fine-tuned a new retrieval algorithm on OpenSearch, replicating key components of the original engine's logic while introducing new functionality where necessary. Through extensive online testing, we demonstrated that the new system performed on par with the original, showing improvements in specific engagement metrics, with negligible effects on revenue.

Figures (3)

• Figure 1: An example sweep of Bayesian optimization
• Figure 2: The overall process of relevance tuning
• Figure 3: Tuning $b$ and $k$ on an example field in our index with all other boosts held constant