Improving Information Retrieval in the Elastic Stack: Introducing ELSER

Source: https://www.elastic.co/search-labs/blog/elastic-learned-sparse-encoder-elser-retrieval-performance
Publisher: Elastic

Summary

Elastic announces ELSER (Elastic Learned Sparse Encoder), their production-grade sparse retrieval model built on SPLADE principles. Reports 17% NDCG@10 improvement over BM25 and describes the distillation-based training approach.

What is ELSER?

ELSER is Elastic’s version of learned sparse retrieval, optimized for the Elasticsearch inverted index infrastructure:

• Based on SPLADE architecture
• 100M parameter BERT-based model
• Produces vocabulary-space sparse vectors (30,522 BERT vocabulary dimensions)
• Deployed natively in Elasticsearch/OpenSearch

Training Methodology: Knowledge Distillation

ELSER was trained using a powerful ensemble teacher:

Teacher: MiniLM (fast bi-encoder) + MonoT5-3B (large cross-encoder reranker)
    ↓
Generate high-quality labels for MS MARCO + domain data
    ↓
Student: 100M parameter BERT MLM model
    ↓
ELSER: compressed, fast sparse encoder

The teacher ensemble captures complementary signals:

• MiniLM: semantic similarity
• MonoT5-3B: nuanced relevance judgment

This is an example of Retrieval Pipeline distillation — compressing a multi-stage pipeline into a single model.

Performance Results

On MS MARCO dev set (NDCG@10):

Model	NDCG@10	Notes
BM25	0.228	Lexical baseline
ELSER v1	0.267	+17% over BM25
Dense bi-encoder	0.253	Semantic only
Hybrid BM25+Dense	0.271	Best combination

ELSER outperforms both BM25 and standard dense bi-encoders individually, approaching hybrid quality with a single model.

ELSER v2 Improvements

ELSER v2 brings:

• Better OOD (out-of-distribution) performance
• Faster inference
• Lower memory footprint
• Improved handling of numeric/code content

Integration in Elasticsearch

# Elasticsearch 8.x API
GET /my_index/_search
{
  "query": {
    "text_expansion": {
      "ml.tokens": {
        "model_id": ".elser_model_2_linux-x86_64",
        "model_text": "find documents about machine learning"
      }
    }
  }
}

Elasticsearch stores ELSER sparse vectors natively in its inverted index — no separate vector database required.

ELSER as Hybrid Component

For best results, Elastic recommends ELSER + BM25 hybrid:

{
  "query": {
    "bool": {
      "should": [
        {"match": {"content": "query"}},  // BM25
        {"text_expansion": {"ml.tokens": {"model_id": ".elser_model_2", "model_text": "query"}}}  // ELSER
      ]
    }
  }
}

Related Articles

• SPLADE for Sparse Vector Search Explained — SPLADE foundations
• SPLADE - Sparse Bi-Encoder BERT Model for First-Stage Ranking — original research
• Hybrid Search SPLADE Sparse Encoder — hybrid implementation
• Distilling Retrieval Pipelines to a Single Embedding Model — related distillation concept

Related Concepts

• ELSER — primary topic
• SPLADE — architecture basis
• Sparse Vector Retrieval — category
• Retrieval Pipeline — distillation training
• Hybrid Search — recommended deployment
• Embedding Fine-tuning — distillation is a form of training