Compressing float embeddings to one bit per dimension, reducing storage by ~32× with surprisingly small quality loss — useful when speed and scale matter more than marginal precision.
Binary vector embeddings are embeddings where each dimension is reduced from a 32-bit float to a single bit — typically by replacing positive values with 1 and non-positive values with 0. A standard 1024-dimensional float embedding occupies around 4KB; its binary equivalent occupies 128 bytes. The storage reduction is approximately 32×, and similarity search over binary vectors using Hamming distance is significantly faster than cosine similarity over float vectors.
At DEJAN we tested binary embeddings against a state-of-the-art embedding model (mxbai-embed-large-v1) on the MTEB STS12 benchmark. The results showed that binary embeddings at full dimensionality retained the large majority of semantic quality, with meaningful degradation only at very aggressive dimension reductions. When combined with Matryoshka Representation Learning, which nests coarse and fine representations inside a single embedding, binary compression at 256 dimensions achieved storage ratios approaching 50:1 — a 235MB float embedding file reduced to under 5MB — while remaining useful for downstream tasks including similarity search, clustering, and intent classification.
Binary embeddings are a lossy compression. For tasks requiring fine-grained semantic distinction — distinguishing near-synonyms, ranking closely related passages — the quality loss may matter. For large-scale filtering, candidate retrieval, or applications where speed and storage constraints dominate, the trade-off is usually favourable. A common production pattern uses binary embeddings for the first-pass retrieval over a large index, then re-ranks the top candidates using full-precision embeddings.
For SEO practitioners working with large URL inventories or query datasets, binary embeddings make vector-based analysis feasible on standard hardware. Clustering millions of search queries or comparing large content corpora becomes tractable without specialist infrastructure. The technique pairs naturally with vector embedding optimisation workflows.
