nanovec

nanovec is an embedded, serverless vector database for Go, built with the philosophy of SQLite. It runs in-process, requires zero configuration, and persists data to a single file.

Features

• Embedded: No external server required. Links directly into your Go binary.
• Fast: Uses HNSW for approximate nearest neighbor search and SIMD-optimized math.
• Memory Efficient: Supports SQ8 Quantization to reduce RAM usage by 4x.
• Reliable: ACID-compliant storage (via bbolt) ensures data safety on crash.
• Simple: Pure Go, no CGO required (unless standard library dependencies change).

Installation

go get github.com/hungpdn/nanovec

Quick Start

package main

import (
    "fmt"
    "log"
    "github.com/hungpdn/nanovec"
)

func main() {
    // Open database
    cfg := nanovec.Config{
        Dimension: 3,
        IndexType: nanovec.IndexTypeHNSW,
    }
    db, _ := nanovec.Open("mydata.db", &cfg)
    defer db.Close()

    // Insert
    _ = db.Insert("vec1", []float32{1.0, 0.0, 0.0}, map[string]any{"tag": "A"})

    // Search
    results, _ := db.Search([]float32{1.0, 0.0, 0.0}, 1, nil)
    fmt.Printf("Found: %s\n", results[0].ID)
}

🚀 Instant Startup with Read-Only Mode (Mmap)

For large datasets (e.g., 10M+ vectors), loading the index into RAM can take minutes. Nanovec supports Memory Mapping (mmap) for FlatIndex, allowing instant startup (0ms load time) and OS-managed memory paging.

cfg := nanovec.Config{
    Dimension: 128,
    IndexType: nanovec.IndexTypeFlat,
    ReadOnly:  true, // <--- Enable Zero-Copy Load
}

// Opens instantly, even with 100GB data!
db, _ := nanovec.Open("large_data.db", &cfg)

Benchmarks

Benchmarks run on an Intel Core i7-8850H CPU @ 2.60GHz using a dataset of 128-dimensional vectors.

🚀 Search Performance (HNSW)

Latencies are measured on a dataset of 10,000 vectors.

Operation	Latency (p99)	Throughput	Note
HNSW Search (k=10)	0.65 ms	~1,900 QPS	Sub-millisecond search

💾 Ingestion Speed (Write Throughput)

Nanovec uses ACID storage (BoltDB). Using InsertBatch significantly reduces fsync overhead.

Method	Time per Item	Operations/Sec	Recommendation
Batch Insert (1k items)	0.06 ms	~15,400 ops/s	✅ Highly Recommended
Sequential Insert	45.9 ms	~22 ops/s	❌ Only for small updates

🧠 Memory vs. Speed (Float32 vs. SQ8)

Micro-benchmarks on Dot Product (1536 dimensions).

Index Type	SIMD Speed (ns/op)	RAM Usage	Best For
Float32	575 ns	100% (6KB/vec)	Maximum Speed
SQ8	1199 ns	25% (1.5KB/vec)	Large Datasets (Saves 75% RAM)

License

MIT License

References

• bbolt
• HNSW Paper
• HNSW in Pinecone
• Faiss