Lexical vs. Learned Sparse Retrieval: BM25 vs. SPLADE at Scale

Introduction

Benchmark Results

Retrieval Quality Metrics Comparison

Analysis of SPLADE vs BM25 Performance Across Datasets

Datasets Where SPLADE Outperforms BM25

Datasets Where BM25 Outperforms SPLADE

Training Data Connection

Comparison of BM25 vs SPLADE Term Representation

Term Processing in BM25 vs SPLADE

Term Count Comparison

Justifying SPLADE's Computational Cost

Performance Impact

Performance Metrics

QPS Comparison

Latency (p50 / p95)

Key Takeaways

When does SPLADE outperform BM25?

When is BM25 a better choice?

When should SPLADE be used?

Limitations and Considerations