Module store 

Vector embedding storage.

This module provides storage for vector embeddings associated with entities.

Overview

Entity-based embedding stores

The VectorStore manages dense embeddings organized into named embedding spaces. The SparseVectorStore manages sparse embeddings for high-dimensional vectors with few non-zero elements (e.g., SPLADE embeddings). The MultiVectorStore manages multi-vector embeddings for ColBERT-style late interaction models.

Each entity can have embeddings in multiple spaces (e.g., text_embedding, image_embedding), and each space has a fixed dimension and distance metric.

Collection vector store (entity-to-vector mapping)

The CollectionVectorStore provides dedicated vector storage separate from entity properties. This enables:

• Storage efficiency: Read entities without loading vector data
• Multiple embeddings per entity: Support text, image, summary embeddings
• Independent operations: Update vectors without touching entities
• Cascade deletion: Delete all vectors when entity is removed

Point collections (Qdrant-style)

The PointStore manages points with multiple named vectors and JSON payloads. Points are organized into collections, and each point can have any combination of dense, sparse, or multi-vectors.

Inverted index for sparse vectors

The InvertedIndex provides efficient top-k similarity search for sparse vectors using posting lists and WAND/DAAT algorithms. This is optimized for SPLADE-style sparse retrieval.

Example (Entity embeddings)

use manifoldb_vector::store::VectorStore;
use manifoldb_vector::types::{Embedding, EmbeddingName, EmbeddingSpace};
use manifoldb_vector::distance::DistanceMetric;
use manifoldb_core::EntityId;

// Create a store with a storage backend
let store = VectorStore::new(engine);

// Create an embedding space
let name = EmbeddingName::new("text_embedding")?;
let space = EmbeddingSpace::new(name.clone(), 384, DistanceMetric::Cosine);
store.create_space(&space)?;

// Store an embedding
let embedding = Embedding::new(vec![0.1; 384])?;
store.put(EntityId::new(1), &name, &embedding)?;

// Retrieve it
let retrieved = store.get(EntityId::new(1), &name)?;

Example (Point collections)

use manifoldb_vector::store::PointStore;
use manifoldb_vector::types::{CollectionName, CollectionSchema, VectorConfig, Payload, NamedVector};
use manifoldb_core::PointId;
use std::collections::HashMap;

// Create a point store
let store = PointStore::new(engine);

// Create a collection with schema
let name = CollectionName::new("documents")?;
let schema = CollectionSchema::new()
    .with_vector("dense", VectorConfig::dense(384))
    .with_vector("sparse", VectorConfig::sparse(30522));
store.create_collection(&name, schema)?;

// Insert a point
let mut payload = Payload::new();
payload.insert("title", "Hello World".into());

let mut vectors = HashMap::new();
vectors.insert("dense".to_string(), NamedVector::Dense(vec![0.1; 384]));
vectors.insert("sparse".to_string(), NamedVector::Sparse(vec![(100, 0.5)]));

store.upsert_point(&name, PointId::new(1), payload, vectors)?;

Structs

CollectionVectorStore

Vector storage for collections.

InvertedIndex

Inverted index for sparse vector similarity search.

InvertedIndexMeta

Inverted index metadata for a vector.

MultiVectorStore

A store for multi-vector embeddings (ColBERT-style).

PointStore

A store for points with multiple named vectors.

PostingEntry

A posting list entry: (point_id, weight).

PostingList

A posting list for a single token.

SearchResult

A search result with point ID and score.

SparseVectorStore

A store for sparse vector embeddings.

VectorStore

A store for vector embeddings.

Enums

ScoringFunction

Scoring function for sparse vector similarity.

Constants

TABLE_COLLECTION_VECTORS

Table name for collection vectors.

Functions

decode_vector_value

Decode a vector value from storage.

encode_vector_value

Encode a vector value for storage.