RETRIEVAL RESEARCH

Sentence Transformers demonstrates training embeddings for compressed multi-vector retrieval

By IDAKIM Digital · Published November 13, 2026

A new walkthrough and code show how pooling-aware objectives can preserve retrieval quality at a smaller vector budget.

Hugging Face published Tom Aarsen’s token-pooling-aware training walkthrough on November 11. Multi-vector retrieval keeps several embeddings per document; merging related token vectors reduces storage but can damage distinctions needed for search. The example trains the same encoder against both unpooled and compressed document representations, using a loss wrapper in Sentence Transformers. In the author’s NanoBEIR experiment, pooling-aware training retains more of the uncompressed retrieval score than ordinary training at the same fourfold vector reduction. This is a reproducible experimental comparison, not a universal retrieval guarantee. Query embeddings remain unpooled in the demonstrated configuration, and adaptive vector budgets need calibration on representative documents. Optional CUDA kernels and training dependencies have their own hardware requirements.

Why it matters

Compare ordinary and pooling-aware encoders on held-out queries from your corpus. Measure index size, relevance and end-to-end latency together before replacing an existing search index.

Source: Hugging Face / Tom Aarsen · November 11, 2026

Product claims are attributed to the linked publisher. IDAKIM Digital has not independently tested these claims.