How to enhance Elasticsearch's relevance algorithm with business logic

Elasticsearch’s default relevance scoring (built on BM25, a refinement of TF-IDF) is genuinely good at one thing: ranking results by how well they match a search query’s text. What it has no concept of is business context — whether a matched product is actually the one worth buying right now.

The pattern: boost, don’t replace

The reliable way to add business logic to Elasticsearch relevance is to leave its native scoring alone and layer a boost function on top of it, most commonly via a function_score query. Elasticsearch computes its normal text-relevance score as usual, and a set of business-defined functions adjust that score up or down before results are returned. This keeps search quality (matching intent) and business priority (what’s worth surfacing) as two separate, independently tunable concerns.

Common signals worth boosting on

In practice, the signals worth factoring into a boost layer tend to be things a generic search engine can’t know on its own:

• Lifecycle stage — new or actively promoted products can be boosted; phase-out products deprioritized.
• Segmentation — boosting listings that match the audience or use case a customer is likely in.
• Stock and market availability — surfacing what a customer can actually buy in their market ahead of what they can’t.
• Business priority or margin — some organizations also boost by strategic importance, though this needs a light touch to avoid feeling manipulative to the buyer.

What makes a boost layer maintainable

A boost layer that’s hardcoded once and never revisited goes stale as fast as a manual sort order would. What’s held up well in my experience: keep the weighting logic in one clearly owned calculation model rather than scattered across queries, make it easy to add a new signal without touching existing ones, and treat any change to it like a product change — hypothesis, before/after comparison, sign-off before it ships. I lean on the same structured, role-tailored UAT approach my squad uses for everything else we ship.

This is the model behind the relevance boost layer I built on endress.com’s onsite search.