Crate nitrite_vector 

Nitrite Vector — HNSW ANN index & RAG store for Nitrite

This crate adds an approximate-nearest-neighbour (ANN) vector index to the Nitrite embedded database, backed by a hand-rolled, persistent HNSW (Hierarchical Navigable Small World) graph, plus a thin RagStore convenience layer for retrieval-augmented-generation workloads.

Embeddings are provided by the caller (bring-your-own vectors); this crate does not generate embeddings.

Quick start (raw collection API)

use nitrite::nitrite::Nitrite;
use nitrite::common::PersistentCollection;
use nitrite_vector::{VectorModule, VectorIndexConfig, vector_index_options, vector_field};
use nitrite_vector::distance::Metric;

let db = Nitrite::builder()
    .load_module(VectorModule::new(VectorIndexConfig::new(3, Metric::Cosine)))
    .open_or_create(None, None)?;

let collection = db.collection("docs")?;
collection.create_index(vec!["embedding"], &vector_index_options())?;

// ... insert documents whose `embedding` field is a numeric array ...

let filter = vector_field("embedding").nearest(vec![0.1, 0.2, 0.3], 5).build();
let results = collection.find(filter)?;

RAG store

use nitrite_vector::RagStore;
use nitrite_vector::distance::Metric;

// `Metric::Cosine` must match the metric configured on the VectorModule.
let store = RagStore::create(&db, "kb", Metric::Cosine)?;
store.add("hello world", embedding, doc!{ "source": "wiki" })?;
let hits = store.search(query_vector, 5).run()?;

Security note

Vectors are stored in plaintext by both backends (embeddings are generally invertible back to content — treat them as the data itself). The DiskANN backend writes its files next to the database, outside whatever the storage adapter provides, and therefore requires a persistent db_path.

Re-exports

pub use diskann::DiskAnnConfig;

pub use diskann::DiskAnnIndex;

pub use distance::Metric;

pub use filter::value_to_vector;

pub use filter::vector_to_value;

pub use filter::VectorNearestFilter;

pub use filter::VECTOR_INDEX;

pub use fluent::vector_field;

pub use fluent::VectorFluentFilter;

pub use fluent::VectorNearestBuilder;

pub use indexer::VectorIndexer;

pub use module::VectorModule;

pub use module::VectorModuleBuilder;

pub use precision::Precision;

pub use rag::RagStore;

pub use rag::SearchHit;

pub use rag::SearchQuery;

pub use vector_index::derive_vector_map_name;

pub use vector_index::HnswBackend;

pub use vector_index::IndexBackend;

pub use vector_index::VectorIndex;

pub use vector_index::VectorIndexConfig;

Modules

diskann

Disk-resident DiskANN backend: a single-layer Vamana graph plus full vectors stored on disk in a memory-mapped flat file (flat_store), with product-quantized codes resident in RAM for fast approximate traversal and exact re-ranking from the on-disk vectors.

distance

Distance metrics for vector similarity search.

filter

Vector search filter.

fluent

Fluent API for building vector search filters.

hnsw

In-memory HNSW (Hierarchical Navigable Small World) graph.

indexer

The vector indexer: implements Nitrite’s NitriteIndexerProvider so an HNSW index can be created, maintained, and queried through the standard collection API.

module

Nitrite module that registers the vector indexer, plus a fluent builder that surfaces every configurable knob.

node

Serializable records for the HNSW graph.

precision

User-selectable stored-vector precision.

rag

A thin retrieval-augmented-generation (RAG) store over a Nitrite collection.

vector_index

Durable HNSW index: bridges the in-memory Hnsw graph with a Nitrite NitriteMap so the graph survives restarts and participates in the store’s atomicity.

Functions

vector_index_options

Creates IndexOptions for a vector index.