Search using LTR - Elastic Docs

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

How to apply a deployed native Elasticsearch LTR model at search time (8.12.0+, subscription-gated). Two ways to use it.

---

As a rescorer

Add a learning_to_rank block under rescore, referencing the model_id and passing named params (e.g. query_text) into the feature query templates. A first-pass query supplies the candidates to rescore.

"rescore": {
  "learning_to_rank": {
    "model_id": "ltr-model",
    "params": { "query_text": "the quick brown fox" }
  },
  "window_size": 100
}

As a retriever (Stack 9.1+)

LTR can also sit in the retriever pipeline via a rescorer retriever that wraps an inner retriever (e.g. a standard retriever) and applies the same learning_to_rank rescore over its results.

Known limitation: window_size

LTR scores are generally not comparable to first-pass query scores and can be lower, so a non-rescored doc could otherwise outrank a rescored one. Therefore window_size is mandatory for LTR rescorers and must be >= from + size. When paginating, keep window_size constant across pages (only from changes) — otherwise top hits shift confusingly as the user steps through pages.

Related Concepts

• Elasticsearch Learning to Rank · Learning to Rank · Reranking · Retrieval Pipeline

Related Articles

• Learning To Rank (LTR) - Elasticsearch Native — the overview / training side
• We’re Bringing Learning to Rank to Elasticsearch — the predecessor OSC plugin

Companies

• Elastic