Antarys Vector Database

Blazingly Fast Vector Database for Everyone

Documentation | Discord | Twitter

---

Antarys is a high-performance vector database engineered for production-scale AI applications. Built from the ground up for speed, it delivers 1.5-2x faster text query and 6-8x faster image queries compared to leading alternatives, while maintaining superior recall accuracy.

🎥 See Antarys In Action With Image Search

Table of Contents

• Performance Benchmarks
  • Write Performance
  • Batch Operations
  • Query Performance
  • Search Quality & Recall
• Installation
  • Download Antarys Database
  • Install Python Client
  • Alternative: Node.js Client
• Quick Start
• Core Features
  • Collections
  • Built-in Text Embeddings
  • Vector Operations
    • Upsert Vectors
    • Query Vectors
    • Delete Vectors
• Performance Optimization
  • Client Configuration
  • Scale-Based Recommendations
  • HNSW Index Tuning
• Advanced Features
  • Dimension Validation
  • Cache Management
  • Health Monitoring
• Type Safety
• Resources
• License

Performance Benchmarks

Benchmarked against leading vector databases using the OpenAI-compatible DBpedia Dataset (1M vectors, 1536 dimensions).

Write Performance

Database	Throughput (vectors/sec)	Performance vs Antarys
Antarys	2,017	Baseline
Chroma	1,234	1.6x slower
Qdrant	892	2.3x slower
Milvus	445	4.5x slower

Batch Operations

Database	Avg Batch Time (ms)	P99 Latency (ms)
Antarys	495.7	570.3
Chroma	810.4	890.2
Qdrant	1,121.6	1,456.8
Milvus	2,247.3	3,102.5

Query Performance

Database	Throughput (queries/sec)	Avg Query Time (ms)	P99 Latency (ms)
Antarys	602.4	1.66	6.9
Chroma	340.1	2.94	14.2
Qdrant	19.4	51.47	186.3
Milvus	4.5	220.46	892.1

Search Quality & Recall

Database	Recall@100 (%)	Standard Deviation
Antarys	98.47%	0.0023
Chroma	97.12%	0.0034
Qdrant	96.83%	0.0041
Milvus	95.67%	0.0056

View full benchmark repository →

Installation

Download Antarys Database

Install Antarys with our one-line installer (macOS ARM and Linux x64):

curl -fsSL http://antarys.ai/start.sh | bash

Install Python Client

pip install antarys

For accelerated performance, install optional dependencies:

pip install antarys[performance]
# or individually
pip install numba lz4

Alternative: Node.js Client

npm install @antarys/client

View Node.js documentation →

Quick Start

importasynciofromantarysimportClientasyncdefmain():
# Initialize clientclient=Client(host="http://localhost:8080")
# Create collectionawaitclient.create_collection(
name="my_vectors",
dimensions=1536,
enable_hnsw=True
)
vectors=client.vector_operations("my_vectors")
# Upsert vectorsawaitvectors.upsert([
{
"id": "doc1",
"values": [0.1] *1536,
"metadata": {"category": "AI", "source": "research"}
},
{
"id": "doc2",
"values": [0.2] *1536,
"metadata": {"category": "ML", "source": "tutorial"}
}
])
# Query similar vectorsresults=awaitvectors.query(
vector=[0.15] *1536,
top_k=5,
include_metadata=True,
filter={"category": "AI"}
)
formatchinresults["matches"]:
print(f"ID: {match['id']}, Score: {match['score']:.4f}")
awaitclient.close()
asyncio.run(main())

Core Features

Collections

Create and manage vector collections with optimized parameters:

# Create collection with HNSW indexingawaitclient.create_collection(
name="documents",
dimensions=1536,
enable_hnsw=True,
shards=16,
m=16,
ef_construction=200
)
# List all collectionscollections=awaitclient.list_collections()
# Get collection detailsinfo=awaitclient.describe_collection("documents")
# Delete collectionawaitclient.delete_collection("documents")

Built-in Text Embeddings

Generate embeddings without external API calls:

# Simple embeddingembedding=awaitclient.embed("Hello, World!")
# Batch embeddingsembeddings=awaitclient.embed([
"First document",
"Second document",
"Third document"
])
# Query-optimized embeddingsquery_emb=awaitclient.embed_query("What is artificial intelligence?")
# Document embeddings with progressdoc_embs=awaitclient.embed_documents(
documents=["Doc 1", "Doc 2", "Doc 3"],
show_progress=True
)
# Text similarity comparisonscore=awaitclient.text_similarity(
"machine learning",
"artificial intelligence"
)

Vector Operations

Upsert Vectors

vectors=client.vector_operations("my_collection")
# Single vector upsertawaitvectors.upsert([
{
"id": "vec1",
"values": [0.1, 0.2, 0.3],
"metadata": {"type": "document", "timestamp": 1234567890}
}
])
# Batch upsert for large-scale operationsbatch= []
foriinrange(10000):
batch.append({
"id": f"vector_{i}",
"values": [random.random() for_inrange(1536)],
"metadata": {"category": f"cat_{i%5}"}
})
awaitvectors.upsert_batch(
batch,
batch_size=5000,
parallel_workers=8,
show_progress=True
)

Query Vectors

# Semantic search with filtersresults=awaitvectors.query(
vector=[0.1] *1536,
top_k=10,
include_metadata=True,
filter={"category": "research"},
threshold=0.7,
use_ann=True
)
# Batch queriesquery_vectors= [[0.1] *1536, [0.2] *1536, [0.3] *1536]
batch_results=awaitvectors.batch_query(
vectors=query_vectors,
top_k=5,
include_metadata=True
)
# Get specific vectorvector_data=awaitvectors.get_vector("vec1")
# Count vectorscount=awaitvectors.count_vectors()

Delete Vectors

# Delete by IDsawaitvectors.delete(["vec1", "vec2", "vec3"])

Performance Optimization

Client Configuration

Configure the client for optimal performance based on your workload:

client=Client(
host="http://localhost:8080",
# Connection poolingconnection_pool_size=100,
# HTTP/2 and compressionuse_http2=True,
compression=True,
# Client-side cachingcache_size=1000,
cache_ttl=300,
# Threadingthread_pool_size=16,
# Reliabilityretry_attempts=5,
timeout=120
)

Scale-Based Recommendations

Small Scale (< 1M vectors)

client=Client(
connection_pool_size=20,
cache_size=500,
thread_pool_size=4
)
# Batch operationsbatch_size=1000parallel_workers=2

Medium Scale (1M - 10M vectors)

client=Client(
connection_pool_size=50,
cache_size=2000,
thread_pool_size=8
)
# Batch operationsbatch_size=3000parallel_workers=4

Large Scale (10M+ vectors)

client=Client(
connection_pool_size=100,
cache_size=5000,
thread_pool_size=16
)
# Batch operationsbatch_size=5000parallel_workers=8

HNSW Index Tuning

Optimize HNSW parameters for your accuracy/speed requirements:

# Collection creationawaitclient.create_collection(
name="optimized",
dimensions=1536,
enable_hnsw=True,
m=16, # Connectivity (16-64 for high recall)ef_construction=200, # Construction quality (200-800)shards=32# Parallel processing
)
# Query-time tuningresults=awaitvectors.query(
vector=query_vector,
ef_search=200, # Search quality (100-800)use_ann=True# Enable HNSW acceleration
)

Advanced Features

Dimension Validation

# Validate vector dimensionsis_valid=awaitvectors.validate_vector_dimensions([0.1] *1536)
# Get collection dimensionsdims=awaitvectors.get_collection_dimensions()

Cache Management

# Get cache statisticsstats=vectors.get_cache_stats()
print(f"Cache hit rate: {stats['hit_rate']:.2%}")
# Clear cachesawaitclient.clear_cache()
awaitvectors.clear_cache()

Health Monitoring

# Check server healthhealth=awaitclient.health()
# Get server informationinfo=awaitclient.info()
# Collection statisticscollection_info=awaitclient.describe_collection("vectors")
print(f"Vector count: {collection_info.get('vector_count', 0)}")

Type Safety

Antarys includes comprehensive type definitions:

fromantarys.typesimportVectorRecord, SearchResult, SearchParams# Type-safe vector recordrecord: VectorRecord= {
"id": "example",
"values": [0.1, 0.2, 0.3],
"metadata": {"key": "value"}
}
# Type-safe search parametersparams=SearchParams(
vector=[0.1] *1536,
top_k=10,
include_metadata=True,
threshold=0.8
)

Resources

• Full Documentation - Complete API reference and guides
• Performance Report - Detailed benchmark analysis
• Benchmark Repository - Reproduce performance tests
• Node.js Client - TypeScript/JavaScript SDK

License

Antarys is released under the MIT License.

Community

• Discord - Get help and discuss features
• GitHub Issues - Report bugs or request features
• Twitter - Follow for updates

---

⭐ Star this repo to help more developers discover Antarys!