> What content **resembles** this? ^question ## Definition > A [[Database|database]] where: > - [[Notes/Index|Index]]: [[Embedding|embeddings]] in high-dimensional space > - [[Query]]: retrieves it by **similarity** rather than exact match. ^definition ## Example ```json { "database": [ 0.5, 0.6, 0.1], "graph": [ 0.1, 0.9, 0 ], "vector": [ 0.9, 0.1, 0.5 ], "graph DB": [ 0.3, 0.7, 0.1 ], "vector DB": [ 0.7, 0.3, 0.3 ], "hybrid DB": [ 0.6, 0.5, 0.5 ] } ``` ^example A query is itself embedded; the database returns the nearest neighbours in vector space. ## Retrieval mechanism Approximate nearest neighbour (ANN) algorithms - HNSW, IVF, product quantisation - trade exactness for speed. Production vector stores return top-k results in milliseconds across millions to billions of vectors. Similarity is implicit in the embedding model. There is no schema for *why* two items are close; only that the model's geometry put them there. ^retrieval ## What it solves - [[Semantic search]] over a document corpus - Recommendation by content or behaviour similarity - Single-document Q&A where the answer fits in one chunk - Bootstrapping a [[Knowledge base]] with no curated schema - Fuzzy matching that tolerates typos and paraphrasing ^what-it-solves ## Limits - [[Opacity]] - no explanation of why two items are similar, only a score - [[Embedding drift]] - swapping the embedding model shifts the entire similarity landscape - Multi-entity queries - accuracy degrades sharply when the answer requires combining facts across entities the embedding model never explicitly linked ^limits