Benchmarking a Korean fine-tuned SPLADE-based Neural Sparse Model and Amazon OpenSearch Hybrid Search | Amazon Web Services

AWS benchmarked a Korean SPLADE v3-style neural sparse encoder on Amazon OpenSearch Service.

The study compared BM25, Titan Embeddings V2, and hybrid retrieval on the MIRACL-ko benchmark.

Results highlight the tradeoffs between lexical, dense, and neural sparse search for Korean queries.

The work offers a practical guide for choosing indexing and ranking strategies in real OpenSearch deployments.