vctrs 0.2.0

A fast embedded vector database. Rust core, Python API. 4-10x faster than ChromaDB.

vctrs

A fast embedded vector database. Rust core with Python and Node bindings.

Vector search as a library, not a service. No server, no config. pip install vctrs and go.

Why vctrs

ChromaDB, Pinecone, and Qdrant are databases you run. vctrs is a library you embed. Use it when:

• You're building a tool or library that needs vector search without inflicting a 200MB dependency tree on users
• You need high-throughput batch search — deduplication, clustering, similarity joins in tight loops
• You want sub-millisecond search without the overhead of client-server round trips
• You're building in Rust and need a native vector search crate

Install

pip install vctrs        # Python
npm install @yang-29/vctrs  # Node

Python

from vctrs import Database

db = Database("./mydb", dim=384, metric="cosine")

# CRUD
db.add("doc1", vector, {"title": "hello"})
db.upsert("doc1", new_vector, {"title": "updated"})
db.update("doc1", metadata={"title": "changed"})   # update metadata only
db.update("doc1", vector=new_vector)                # update vector only
db.delete("doc1")
db.get("doc1")                          # → (vector, metadata)
"doc1" in db                            # → True

# Batch insert (parallel HNSW construction)
db.add_many(ids, vectors, metadatas)

# Search
results = db.search(query_vector, k=10)  # → [SearchResult(id, distance, metadata), ...]
batch = db.search_many(query_vectors, k=10)  # parallel multi-query search

# Filtered search
results = db.search(query, k=10, where_filter={"category": "science"})
results = db.search(query, k=10, where_filter={"category": {"$ne": "sports"}})
results = db.search(query, k=10, where_filter={"category": {"$in": ["sci", "tech"]}})
results = db.search(query, k=10, where_filter={"score": {"$gte": 0.5, "$lt": 0.9}})

# Maintenance
db.compact()                # reclaim deleted vector slots
db.enable_quantized_search()  # SQ8 quantized HNSW traversal + f32 re-ranking
db.save()                   # persist to disk

# Diagnostics
stats = db.stats()          # → dict with graph metrics, memory usage, etc.

# Context manager (auto-save on exit)
with Database("./mydb", dim=384) as db:
    db.add("doc1", vector)

Options: m=16 (HNSW links), ef_construction=200 (build quality), quantize=True (SQ8, ~4x smaller on disk).

Node

const { VctrsDatabase } = require("@yang-29/vctrs");

const db = new VctrsDatabase("./mydb", 384, "cosine");

// CRUD
db.add("doc1", vector, { title: "hello" });
db.upsert("doc1", newVector, { title: "updated" });
db.update("doc1", null, { title: "changed" }); // metadata only
db.delete("doc1");
db.get("doc1"); // → { vector, metadata }
db.contains("doc1"); // → true

// Batch
db.addMany(ids, vectors, metadatas);

// Search
const results = db.search(queryVector, 10); // → [{ id, distance, metadata }, ...]
const batch = db.searchMany(queryVectors, 10); // parallel multi-query search

// Filtered search
db.search(query, 10, null, { category: "science" });
db.search(query, 10, null, { score: { $gte: 0.5, $lt: 0.9 } });

// Maintenance
db.compact();
db.enableQuantizedSearch();
db.save();

// Diagnostics
const stats = db.stats(); // → { numVectors, numDeleted, avgDegreeLayer0, ... }

Metrics: "cosine" (default), "euclidean" / "l2", "dot" / "dot_product".

Filter operators

Operator	Description	Example
$eq (default)	Equals	{ field: "value" }
$ne	Not equals	{ field: { $ne: "value" } }
$in	In list	{ field: { $in: ["a", "b"] } }
$gt	Greater than	{ field: { $gt: 10 } }
$gte	Greater than or equal	{ field: { $gte: 10 } }
$lt	Less than	{ field: { $lt: 20 } }
$lte	Less than or equal	{ field: { $lte: 20 } }

Multiple keys in a filter object are ANDed together.

Performance

10,000 vectors, 384 dimensions, cosine, Apple M-series:

	vctrs	ChromaDB	numpy
Search k=10	0.14ms	0.93ms	0.16ms
Insert 10k	518ms	2367ms	—
Load from disk	1.2ms	1.6ms	—

6-7x faster than ChromaDB. Matches raw numpy. Uses Apple Accelerate / OpenBLAS for batch distance computation, mmap for instant loads.

How it works

• HNSW index with flat contiguous vector storage for cache locality
• Optional scalar quantization (SQ8) for ~4x smaller on-disk storage with full-precision re-ranking
• In-graph filtered search (no over-fetch retry loop)
• Auto brute-force for small datasets (100% recall, BLAS-accelerated)
• Memory-mapped vectors — zero-copy load, OS-managed paging
• GC/compaction to reclaim soft-deleted vector slots
• SimSIMD for per-vector SIMD (ARM NEON, x86 AVX2/512)
• Rayon for parallel index construction and batch search
• PyO3 + maturin for zero-copy Python/numpy bindings

Rust

[dependencies]
vctrs-core = "0.2"

use vctrs_core::db::{Database, Filter, HnswConfig};
use vctrs_core::distance::Metric;

let db = Database::open_or_create("./mydb", 384, Metric::Cosine)?;
db.add("doc1", embedding, Some(json!({"title": "hello"})))?;

// Search
let results = db.search(&query, 10, None, None)?;

// Filtered search
let filter = Filter::Gte("score".into(), 0.5);
let results = db.search(&query, 10, None, Some(&filter))?;

// Batch search (parallel)
let batch = db.search_many(&[&q1, &q2], 10, None, None)?;

// Maintenance
db.compact()?;
db.enable_quantized_search();
let stats = db.stats();

// Custom HNSW config + quantization
let config = HnswConfig { m: 32, ef_construction: 400, quantize: true };
let db = Database::open_or_create_with_config("./mydb", 384, Metric::Cosine, config)?;

Errors are typed via VctrsError enum (DimensionMismatch, DuplicateId, NotFound, Io, etc.).

Examples

See examples/ for complete applications:

• Semantic Search — search over documents with sentence embeddings
• Deduplication — find near-duplicate items in a dataset
• RAG — retrieval-augmented generation with local LLM

Build from source

git clone https://github.com/yang-29/vctrs.git && cd vctrs
python -m venv .venv && source .venv/bin/activate
pip install numpy maturin
maturin develop --release

License

MIT