ELSER
Definition
ELSER (Elastic Learned Sparse EncodeR) is Elastic’s pre-trained learned sparse retrieval model, built on the SPLADE architecture and integrated natively into Elasticsearch as a text_expansion query type. It achieves strong zero-shot retrieval performance without requiring task-specific fine-tuning.
Developed by Thomas Veasey and Quentin Herreros at Elastic.
How It Differs from SPLADE
ELSER makes several training innovations beyond standard SPLADE:
Improved teacher model: Uses an ensemble of MiniLM and MonoT5-3B cross-encoders with monotonic score transformation to smooth distribution skew.
Training insight: Vocabulary tokens function as interdependent vector components, not independent lexical items — removing low-scoring tokens reduces quality.
FLOPS regularizer behavior: 99% of token pruning occurs in the first 50,000 batches (warmup phase), acting as implicit stop-word removal.
Performance
BEIR benchmark vs. BM25: 10 wins, 1 draw, 1 loss across 12 datasets — 17% average NDCG@10 improvement.
Model size: only 100M parameters (much smaller than many generative models).
Key Technical Properties
- Storage efficient: Documents expand to ~100 tokens on average (approximate size parity with text indices)
- Inverted index compatible: Leverages Lucene — no special ANN infrastructure needed
- Interpretable: Token highlighting shows which terms drove matches
- Latency-quality tunable: FLOPS regularizer controls this tradeoff
Integration
Available in Elasticsearch as text_expansion query clause — one-click deployment without tuning.
Related Concepts
-
Embeddings — parent concept
-
Sparse Embeddings — ELSER is Elastic’s production learned sparse embedding model
-
SPLADE — the architecture ELSER is based on
-
Sparse Vector Retrieval — ELSER is a learned sparse model
-
Hybrid Search — ELSER + BM25 (or dense) hybrid typically outperforms either alone
-
Cross-Encoder — used as teacher in ELSER’s distillation training