Elasticsearch Learning to Rank

In-engine LTR: running a learned ranking model inside Elasticsearch as a second-stage rescorer, rather than in an external service. The model is trained offline (a LambdaMART / GBDT), uploaded into the cluster, and applied to rescore the top-N hits of a cheap first-pass query. Features are computed at query time from templated Query DSL queries — the relevance score of match/multi_match etc. becomes a feature value.

Elasticsearch offers in-engine LTR: the model runs inside the cluster as a rescore phase, not as a separate service. There are two distinct modules — the third-party OpenSource Connections plugin and Elastic’s native LTR (8.12+) — and this note focuses on that ES-specific distinction.

The broader architectural question — in-engine LTR vs. an external, engine-agnostic secondary re-ranker like Metarank — is not Elasticsearch-specific and is covered in Learning to Rank. Both modules below are in-engine options.

Two Elasticsearch LTR modules

There is no single “Elasticsearch LTR.” Two distinct implementations exist, separated by ~7 years.

1. OpenSource Connections plugin (2017)

The original o19s/elasticsearch-learning-to-rank plugin from OpenSource Connections — for years the only way to do LTR in Elasticsearch, and still in wide use (it powered Wikipedia search).

• Integrates RankLib (the RankyMcRankFace o19s fork): trains LambdaMART/RankNet from a judgment file.
• Ships a custom Elasticsearch script language (ranklib) to store the model, and a custom ltr query that takes a list of Query DSL queries as features plus a model name.
• Applied through rescore over the top-N (running the ltr query directly over a whole collection is too slow).
• Community-maintained; must be installed as a plugin.
• See We’re Bringing Learning to Rank to Elasticsearch.

2. Native Elastic LTR (8.12+)

Elastic shipped first-party LTR in 8.12 (2024), removing the need for the third-party plugin (subscription-gated).

• Model trained outside Elasticsearch as a GBDT — Elastic recommends XGBoost’s XGBRanker (a LambdaMART implementation); inference runs in-engine.
• Models uploaded with eland (Elastic’s Python helper).
• Features extracted via templated query_extractor queries (e.g. title_bm25), grouped into a feature set; three feature categories — document, query, and query-document.
• Served as a learning_to_rank rescorer (8.12+), or as a rescorer retriever (9.1+) inside the retriever pipeline.
• window_size is mandatory for the rescorer (LTR scores aren’t comparable to first-pass scores) and must stay constant across paginated requests.
• See Learning To Rank (LTR) - Elasticsearch Native and Search using LTR - Elastic Docs.

Both modules share the same shape — train a GBDT offline, store it, rescore the top-N with query-derived features — and differ in training library (RankLib vs XGBoost), packaging (plugin vs built-in), and support model (community vs Elastic subscription).

Related Concepts

• Learning to Rank — the general technique this specializes
• LambdaMART — the model both modules serve
• Reranking — LTR is the rescoring stage
• Retrieval Pipeline — first-pass retrieval → LTR rescore
• Feature Store — what external re-rankers add and in-engine LTR avoids
• Judgment Lists · Implicit Judgments — training data
• Hybrid Search — multi-retriever fusion, where external LTR has an edge

Related Tools

• Elasticsearch · Metarank · RankLib · XGBoost · eland · OpenSearch

Articles

• We’re Bringing Learning to Rank to Elasticsearch — OpenSource Connections; the 2017 RankLib-based plugin
• Learning To Rank (LTR) - Elasticsearch Native — Elastic docs; native 8.12+ overview
• Search using LTR - Elastic Docs — Elastic docs; rescorer & retriever usage
• Learn-to-Rank with OpenSearch and Metarank — external re-ranker contrast
• Hybrid Search and Learning-to-Rank with Metarank — multi-retriever fusion case

Companies

• OpenSource Connections — authored the original plugin
• Elastic — ships native LTR