Document Similarity Matching
Which documents are like this one? — the primitive behind dedup, precedent search, and fraud rings.
Document similarity matching is the measurement of how alike documents are — a primitive that, pointed in different directions, becomes a dozen capabilities: deduplication (are these the same?), version detection (is this a revision of that?), template identification (do these share a form?), precedent retrieval (which past contracts resemble this draft?), clustering (what natural groups does this corpus contain?), and fraud analytics (which "independent" submissions share a suspicious family resemblance?). The question "similar how?" is the design decision: content similarity, structural similarity, and visual similarity are different measurements that different applications need.
The measurement toolkit maps to those senses. Lexical methods (shingling, MinHash, TF-IDF cosine) capture verbatim and near-verbatim overlap — the right tool for duplicates and revisions, robust and explainable. Embedding similarity captures semantic likeness independent of wording — the tool for precedent search and topical clustering, where two contracts can be siblings without sharing sentences. Visual and layout similarity (perceptual hashes, layout fingerprints) captures template kinship — the same form filled differently, which content measures score as dissimilar and template analytics need scored as nearly identical. Production systems often compute several and let the application weight them; fraud detection in particular reads the disagreement between measures (visually near-identical, content-different: a template reused across applicants).
Scale is the engineering constraint — pairwise comparison across millions of documents is quadratic — so similarity systems lean on indexing: locality-sensitive hashing and approximate nearest-neighbor search retrieve candidates cheaply, with precise scoring reserved for the shortlist. Thresholds, as ever, are application-relative and empirically set: the similarity that means "duplicate, merge silently" is far above the one that means "possibly related, show a human," and conflating those two numbers is how similarity systems either miss their targets or bury their users.
The same invoice, submitted three times — dedup catches the copy before it's paid twice.
Meaning as coordinates — documents mapped into vector space where similar content sits close together.
What changed between draft four and draft five — found by machine, ranked by consequence.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo