vecdb

An embeddable vector database in Rust. No server, no daemon, no docker — cargo add, point it at a file, query in-process.

Think SQLite, for vector search. vecdb is a library you link into your app, not a service you run beside it. It stores float32 vectors + JSON metadata on disk and does dense (HNSW), sparse (BM25), hybrid, and filtered search — all embedded.

---

Quickstart

cargo add vecdb-core serde_json

use vecdb_core::Db;
use serde_json::json;

fn main() -> Result<(), vecdb_core::VecDbError> {
    let db = Db::open("data.vecdb")?;              // a directory; created if absent
    let mut docs = db.open_or_create("docs", 4)?; // name, dimension

    docs.insert("a", vec![0.1, 0.2, 0.3, 0.4], json!({ "title": "hello" }))?;
    docs.insert("b", vec![0.9, 0.8, 0.7, 0.6], json!({ "title": "world" }))?;

    let hits = docs.query(&[0.1, 0.2, 0.3, 0.4], 5)?; // top-5 nearest
    println!("nearest: {}", hits[0].id);

    docs.flush()?; // durable via WAL already; flush persists the index for fast reopen
    Ok(())
}

That's the whole thing — no process to start. cargo run and you have vector search.

More: query_text (BM25), query_hybrid (dense+sparse), query_filtered (predicate pushdown on metadata), create_collection_quantized (int8, ~4× smaller index). See the docs.

---

Scale expectations — read this before you load a million vectors

vecdb is built around HNSW, whose index build is the slow part. Query is fast; building the graph is not, and it runs single-node on your machine:

Dataset size	Build time (default ef_construction=200)	Feel
up to ~100k	seconds to a few minutes	instant, use it freely
~1M	~2.5 hours on a laptop	budget for it; build once, reuse

If you're at 1M+ and don't need maximum recall, set ef_construction=100 — ~31% faster build (~1.7h at 1M) for a negligible recall difference (0.9832 → 0.9803 on SIFT1M):

use vecdb_core::{Db, types::CollectionConfig};

let db = Db::open("data.vecdb")?;
let mut cfg = CollectionConfig::new("docs", 128);
cfg.hnsw_ef_construction = 100;
let docs = db.create_collection_with(cfg)?;

Query latency stays competitive with server databases at 1M (in-process, no HTTP hop). The honest full numbers vs qdrant / chroma / pgvector — including where vecdb loses — are in BENCHMARK.md. Bulk-loading? Use Collection::insert_batch — it builds the index once instead of on every insert.

---

Features

• Dense search — HNSW (via instant-distance), cosine / euclidean / dot metrics, SIMD distance kernels
• Sparse search — built-in BM25 inverted index, no external search engine
• Hybrid — weighted-sum or reciprocal-rank fusion of dense + sparse
• Filtered search — predicate pushdown on JSON metadata; declare indexed_payload_fields for fast selective filters
• Quantization — opt-in int8 (~4× smaller index) or 1-bit binary (~32×) per collection
• Durable — write-ahead log, crash recovery, reopen persists
• Embedded — zero network dependencies in vecdb-core; SQLite (bundled) for metadata

---

Server mode (optional)

Prefer a REST service? The vecdb-api crate wraps the embedded core in an axum HTTP server (Prometheus metrics, API key auth, Python/TypeScript SDKs). It's secondary to the embedded story — most users want the library. See crates/vecdb-api.

---

License

MIT — see LICENSE.