Glossary

The vector search vocabulary memista uses, defined in plain language.

Vector search: Finding items whose embeddings are closest to a query embedding, rather than matching keywords. The basis of semantic search, RAG retrieval, and "more like this".
Embedding: A fixed-length vector of floats produced by a model to represent text (or other data). Similar meanings produce nearby vectors. memista stores and searches these.
HNSW: Hierarchical Navigable Small World — a graph-based approximate nearest-neighbour index. It trades exactness for speed and is the algorithm USearch uses under the hood.
USearch: The vector index library memista is built on (version 2.19.x), with SIMD acceleration via simsimd. It provides the HNSW index and the .usearch on-disk format.
ANN (approximate nearest neighbour): Search that returns very-close neighbours quickly without guaranteeing the exact closest. HNSW is an ANN method; a flat/brute-force index is exact but slower.
Inner product (IP): The distance metric memista uses (MetricKind::IP). For normalised embeddings, ranking by inner product is equivalent to ranking by cosine similarity.
Quantization: How vector components are stored numerically. memista uses F32 (ScalarKind::F32) — full 32-bit floats — for the index.
Partition: An isolated dataset identified by a database_id. Each partition has its own SQLite table (chunks_<database_id>) and its own <database_id>.usearch index file.
Chunk: The unit memista stores: an embedding, the text it represents, and a metadata string. Insert chunks; search returns ranked chunks.
Recall: The fraction of the true nearest neighbours an approximate index actually returns. HNSW recall is tunable via connectivity and expansion parameters (fork the index helper to change them).
WAL (write-ahead logging): The SQLite journal mode memista uses for the metadata database — it allows concurrent reads during writes and makes crash recovery cleaner.