Jaccard Similarity for sparse vectors.

Set-like and sparse use cases

Jaccard is natural for set-like or bag-of-words data: document term sets, user interaction sets, or any sparse representation where presence/absence matters more than magnitude. The corresponding distance is often 1 − J (Jaccard distance).

It is not a metric in the strict sense when used on weighted vectors (e.g. TF-IDF) unless you binarize; for weighted sparse vectors, cosine or overlap coefficients are more common. Practical tip: use Jaccard when you care only about which dimensions are “on”; when weights matter (e.g. term importance), use cosine similarity instead.

Role in vector databases

In vector databases, Jaccard is less common than cosine or L2 for dense embeddings, but it appears in sparse retrieval, duplicate detection, and hybrid setups where one signal is set-based.

MinHash and LSH can approximate Jaccard quickly for large sets, enabling scalable set similarity search. Trade-off: exact Jaccard is O(|A|+|B|) for set size; MinHash gives bounded error with sublinear time. Locally sensitive hashing (LSH) for VDBs describes how LSH is used in vector and set search.

Practical tip: when combining set-based signals with dense vector search (e.g. hybrid search), use Jaccard or overlap on keyword/tag sets and cosine or L2 on embeddings; then fuse the two rankings (e.g. RRF or weighted combination). Metadata filtering can further restrict to a subset before or after Jaccard similarity.

When to choose Jaccard vs. cosine

Choose Jaccard when only presence/absence of dimensions (or tokens) matters—e.g. binary features, tag sets, or binarized bag-of-words. Choose cosine when vector components have weights (e.g. TF-IDF, embeddings); cosine uses magnitude, Jaccard ignores it. For hybrid search combining vector and keyword signals, Jaccard or overlap on keyword sets can be one of the signals alongside dense vector similarity.

Pipeline summary: represent documents or items as sets (e.g. token IDs or dimension indices), build an index that supports set or sparse similarity (or use MinHash signatures), and query with the same set representation. Not all vector DBs support Jaccard natively; some expose it for sparse or metadata-only use cases, or you compute Jaccard in application code after filtering by other criteria.