How it works

memista pairs two well-known pieces — USearch for the index and SQLite for the metadata — and keeps them in sync with a shared chunk id. An Actix-web layer exposes three endpoints on top.

┌─────────────────────────────────────────────────────────┐
│  POST /v1/insert    POST /v1/search    DELETE /v1/drop  │
│             (actix-web + apistos OpenAPI)               │
├─────────────────────────────────────────────────────────┤
│                                                         │
│   ┌──────────────────┐         ┌─────────────────────┐  │
│   │   SQLite (WAL)   │◀───────▶│   USearch index     │  │
│   │ chunks_<db_id>   │  chunk  │  <db_id>.usearch    │  │
│   │ (text, metadata) │   id    │  HNSW · IP · f32    │  │
│   └──────────────────┘         └─────────────────────┘  │
│                                                         │
└─────────────────────────────────────────────────────────┘

The insert path

On POST /v1/insert, the row (text + metadata) is written to the SQLite table chunks_<database_id> first. The returned chunk_id becomes the USearch key for the vector. Because the SQLite id and the USearch key are the same integer, the two stores never drift. After the batch, the <database_id>.usearch index file is saved to disk.

The search path

On POST /v1/search, USearch returns ranked keys for the query vector, ordered by inner-product distance (MetricKind::IP, F32). memista then hydrates the text and metadata for those keys back from SQLite and returns them with their distances. One index lookup, one metadata read.

Partitions

Each database_id is a partition: its own SQLite table and its own .usearch file. Isolation is by name. DELETE /v1/drop removes both. See persistence and the .usearch file.

Defaults you can tune

The stock build passes USearch's default connectivity, expansion_add, and expansion_search (all 0, meaning "use USearch defaults"). To trade recall against latency, fork the load_or_create_index helper. Embedding dimensions are likewise fixed at index creation — see changing embedding dimensions.