Sentence Transformers SparseEncoder: Sparse Retrieval in a Few Lines
tomaarsen · x · 2026-09-17
Third post in tomaarsen's SparseUp thread: Sentence Transformers now ships a SparseEncoder interface, fully compatible with the new Linkup-sparseup-embed-v1.
- Sparse encoders: mostly-zero vectors enable efficient large-scale retrieval; non-zero dims map to specific tokens for interpretability; combinable with dense embeddings for hybrid search
- Usage: load SparseEncoder("naver/splade-cocondenser-ensembledistil"), call encode() for [3, 30522] sparse vectors, similarity() uses dot product by default
- sparsity() reports sparsity ratio (>99% typical) and average active dims; promptname / encodequery() / encodedocument() handle query vs document prompts
Full docs on the Sentence Transformers site.
More from Models
- Claim: Kimi-K3 is 'chronically undertrained,' casting doubt on Moonshot's training budget — scaling01 · 2026-09-17
- TypeSafe's Jev ditches text generation for instant, calibrated numerical answers — and it plays Doom at 7 req/s — JnBrymn · 2026-09-17
- GPT Image 2.5 excels at unblurring images, eating yet another niche API — shekitup · 2026-09-17
- Papers with Code Launches MCP Server; Claude Code Used to Infer Jev's Architecture — NielsRogge · 2026-09-17
- Jev passes 8/9 computer-use tasks, makes decisions 13.6x faster than Astra/Codex — iamrobotbear · 2026-09-17
- Claude Max users can now buy usage resets for $40, signaling end of free resets — MrBobrowitz · 2026-09-17