More Like This (MLT)

More Like This, abbreviated MLT, is a search technique that returns documents similar to a given example document, rather than documents matching an explicit textual query. The user starts with a known relevant item — a paragraph, a paper, a product description — and the system surfaces other items that are likely relevant by the same measure. The technique predates modern machine learning and remains in active use across search engines, content platforms, and legal-discovery tools as a low-friction way to navigate a corpus.

Mechanically, classical MLT extracts the most discriminative terms from the source document and runs them as a search query against the index, often with term-frequency–inverse-document-frequency weighting to prefer terms that characterize the source rather than the whole corpus. The retrieved documents are then re-ranked by similarity to the original. Modern vector-based variants skip the term-extraction step: the source document is embedded into a vector space and the system retrieves its approximate nearest neighbors. The two approaches compose — vector retrieval followed by a re-ranker that blends in lexical signals is the dominant production pattern.

MLT shines when the user cannot articulate a query but can point at an example, which is common in research, journalism, and recommendation. The cost is controllability: unlike a keyword query, the user cannot steer the result by adjusting terms. Classical MLT is also sensitive to document length, since longer documents contain more terms and dilute the signal; modern embedding-based variants inherit the biases of the embedding model instead. Hybrid pipelines, in which MLT narrows the candidate set and a downstream re-ranker decides, are the usual way to recover precision lost in the retrieval step.

Open questions include how to expose MLT's behavior to end users in a transparent way — the black-box nature of embedding similarity makes it hard to explain why a given neighbor was returned — and how to compose MLT with filters on metadata (date ranges, authors, sources) without breaking similarity semantics. The boundary between MLT and standard nearest-neighbor search has also blurred: in vector databases, every kNN query is in effect a More Like This query keyed by a document vector rather than an ad-hoc query string.