Crate recern_vector 

Recern Vector — a single-file, embedded, inspectable vector database.

A snapshot and a WAL hold any number of collections. Each collection stores vectors of a fixed dimension, optional JSON metadata, and an HNSW index whose internals can be inspected through Collection::stats, Collection::explain and Collection::estimate_recall.

use recern_vector::{CollectionConfig, Database, Metric, SearchOptions};

let mut db = Database::open_or_create("docs.rvec")?;
let docs = db.create_collection("docs", CollectionConfig::new(3, Metric::Cosine))?;
docs.upsert("a", &[0.1, 0.9, 0.0], None)?;
docs.upsert("b", &[0.8, 0.1, 0.1], None)?;
let hits = docs.search(&[0.2, 0.8, 0.0], 1, &SearchOptions::default())?;
assert_eq!(hits[0].id, "a");
db.save()?;

Structs

Collection

CollectionConfig

CollectionStats

Database

An in-memory database backed by a snapshot and a checksummed .wal journal. Mutations become durable at save(). Use checkpoint() before copying a standalone file. Concurrent handles can read; stale writers fail on save.

HnswParams

HNSW index parameters.

RecallOptions

RecallPoint

RecallReport

Record

A stored record. For cosine collections vector is the normalized vector.

RecordPage

A bounded page in physical insertion order. Cursors are valid only while the collection is unchanged; take a read-only snapshot for an export.

SearchHit

SearchOptions

SearchReport

Search results together with the work it took to produce them.

Enums

Error

Filter

Metadata predicate applied during search.

Metric

Distance function of a collection. For every metric, smaller is closer.

Quantization

Storage precision. Int8 uses a symmetric scale per vector; original floats are not retained. Exact searches are exact over the stored approximation.

Strategy

How a query was answered.

Constants

FORMAT_VERSION

Type Aliases

Result