Learning To Rank (LTR) - Elasticsearch Native

Source: https://www.elastic.co/docs/solutions/search/ranking/learning-to-rank-ltr Author: Elastic (official docs)

Overview page for Elasticsearch’s native LTR, introduced in version 8.12.0 (subscription-gated). The first-party successor to the OpenSource Connections plugin — see Elasticsearch Learning to Rank for the two-module comparison.

---

What it covers

LTR uses a trained ML model as a second-stage re-ranker over a simpler first-stage retrieval. The LTR function takes a list of documents plus a search context (at minimum the user’s query terms; optionally user demographics/geo/age, query length, or per-document signals like the title-field score).

Judgment list

Trained on a judgment list — query/document pairs with relevance grades (binary, or graded 0–4). Can be human- or machine-generated (often from behavioral analytics with human moderation). Quality guidance: balance examples across query types (title vs actor vs director searches) to avoid overfitting, and balance positive vs negative examples.

Feature extraction

Three feature categories:

• Document features — from document properties (e.g. product price).
• Query features — from the query (e.g. number of words).
• Query-document features — document in the context of the query (e.g. BM25 score of the title field).

In Elasticsearch, features are extracted with templated query_extractor queries (e.g. title_bm25 → match on title with {{query}}), used both when building the training set and at inference time.

Models

The objective is to optimally rank against the judgment list under a metric like NDCG or MAP. Elasticsearch relies on gradient-boosted decision tree (GBDT) models for inference. Training happens outside Elasticsearch — LambdaMART is the recommended approach for strong ranking with low inference latency; XGBoost’s XGBRanker is the suggested LambdaMART implementation, and eland helpers integrate the trained model into Elasticsearch.

Related Concepts

• Elasticsearch Learning to Rank · Learning to Rank · LambdaMART · XGBoost · eland · NDCG · Judgment Lists · Reranking

Related Articles

• Search using LTR - Elastic Docs — how to apply the deployed model (rescorer / retriever)
• We’re Bringing Learning to Rank to Elasticsearch — the predecessor OSC plugin
• Learn-to-Rank with OpenSearch and Metarank — external-re-ranker contrast

Companies

• Elastic