mnem-sparse-providers

Learned-sparse encoder adapters for mnem (SPLADE, BGE-M3-sparse, opensearch-doc-v3-distill). Sync, TLS-via-rustls, tokio-free.

Implements mnem_core::sparse::SparseEncoder for three backends per . The sidecar transport (always available) POSTs to a local Python service running the reference SPLADE / BGE-M3 implementation, which keeps the install light and defers weights + tokenization to whatever Python ML infra the user already runs. The ONNX backend (feature onnx) runs inference in-process via ort + tokenizers; fastest, but pulls a heavy dep tree (onnxruntime + SIMD tokenizers) and so is feature-gated to keep WASM targets clean. The mock backend re-exports mnem_core::sparse::MockSparseEncoder for tests and dry-run benchmarks. Learned-sparse is the primary sparse lane after retired the in-tree BM25 lane.

Workspace top: ../../README.md.

Licensed under Apache-2.0.