Getting Started
Turn any Data into Intelligent Data. AI embeddings and Large Language Models depend on your vector data and we are transforming the paradigm of AI data industry
Core Architecture
Micro LLM & Kernel Registry
Vector Database & Data Infra
NATIVE AI
Vectorisation & Embedding Platform
Antarys Compute for Private Data
AHNSW Algorithm
Unlike tradtional sequential HNSW, we are using a different asynchronous approach to HNSW and eliminating thread locks with the help of architectural fine tuning. We will soon release more technical details on the Async HNSW algorithmic approach.
Binary downloads
For Apple ARM
Download for Apple M chips only, intel builds coming soon
For Linux x64
Universal binary for linux for x64 architecture, x86 and ARM coming soon
For Windows x64
Universal binary for windows platforms for both x86, x64 and ARM coming soon
After downloading the antarys binary for your respective platform, you can get started with the following command
./antarys --port=8080For more advance configuration you can pass the following parameters
--port=8080 #API server port
--data-dir=./data #Path to data directory
--enable-gpu=false #Enable GPU acceleration
--shards=16 #Number of data shards
--query-threads=NUM_CPUx2 #Number of query threads
--commit-interval=60 #Auto-commit interval in seconds
--cache-size=10000 #Size of result cache
--enable-hnsw=true #Enable HNSW indexing
--enable-pq=false #Enable Product Quantization
--optimization=3 #Optimization level (0-3)
--max-results=100 #Default maximum results
--batch-size=5000 #Default batch size
--similarity=0.7 #Default similarity threshold
--meta-dir=./metadata #Path to metadata storageClients
For now we have only released a python client which can be installed via
pip install antarysPython Client
Explore python samples, API and fine-tuning from our official docs!
our python client uses asyncio heavily for performance and handling vector operations with threading turned on by default
import asyncio
from antarys import Client
async def main():
# Initialize client with performance optimizations
client = Client(
host="http://localhost:8080",
connection_pool_size=100, # Auto-sized based on CPU count
use_http2=True,
cache_size=1000,
thread_pool_size=16
)
# Create collection
await client.create_collection(
name="my_vectors",
dimensions=1536,
enable_hnsw=True,
shards=16
)
vectors = client.vector_operations("my_vectors")
# Upsert vectors
await vectors.upsert([
{"id": "1", "values": [0.1] * 1536, "metadata": {"category": "A"}},
{"id": "2", "values": [0.2] * 1536, "metadata": {"category": "B"}}
])
# Query similar vectors
results = await vectors.query(
vector=[0.1] * 1536,
top_k=10,
include_metadata=True
)
await client.close()
asyncio.run(main())the python client needs
http2to work, install viapip3 install httpx[http2]
We are working to release gRPC support to our datbase and python client as the immediate feature!
We are extensively working to bring client SDKS for the following platforms
- Typescript
- Java/Kotlin
- .NET
- C/C++
- rust
Benchmarks
For the benchmarks shown below, we used this command
./antarys \
--port=8080 \
--data-dir=./data \
--enable-gpu=false \
--shards=16 \
--query-threads=8 \
--commit-interval=60 \
--cache-size=10000 \
--enable-hnsw=true \
--optimization=3 \
--meta-dir=./metadata \
--enable-pq=trueSee Extended Benchmarks
We have built our benchmarking tool from the ground up to simulate real life performance impacts, check it out now
The following result is generated using our python client with the help of dbpedia 1M vector dataset from huggingface, the benchmarking script can be found here
WRITE PERFORMANCE (batch size 1000)
- Total points inserted: 100000
- Successful batches: 100
- Average batch time: 0.4957s
- Throughput: 2017.54 vectors/sec
- Percentiles: P50=0.4910s, P90=0.5208s, P99=0.5703s
READ PERFORMANCE (100 queries)
- Successful queries: 100
- Average query time: 0.0014s
- Throughput: 692.61 queries/sec
- Percentiles: P50=0.0013s, P90=0.0014s, P99=0.0069s
READ PERFORMANCE (1000 queries)
- Successful queries: 1000
- Average query time: 0.0025s
- Throughput: 395.38 queries/sec
- Percentiles: P50=0.0000s, P90=0.0000s, P99=0.0001s
SUMMARY
======================================================================
Best write performance: 2017.54 vectors/sec (batch size 1000)
Best read performance: 692.61 queries/sec (100 queries)For a similar test we are recording the following performance for Qdrant
WRITE PERFORMANCE (batch size 1000)
- Total points inserted: 100000
- Successful batches: 100
- Average batch time: 0.9631s
- Throughput: 1038.31 vectors/sec
- Percentiles: P50=0.9542s, P90=1.0083s, P99=1.0496s
READ PERFORMANCE (100 queries)
- Successful queries: 100
- Average query time: 0.0417s
- Throughput: 23.99 queries/sec
- Percentiles: P50=0.0325s, P90=0.0559s, P99=0.0916s
READ PERFORMANCE (1000 queries)
- Successful queries: 1000
- Average query time: 0.0359s
- Throughput: 27.88 queries/sec
- Percentiles: P50=0.0328s, P90=0.0550s, P99=0.0564s
SUMMARY
======================================================================
Best write performance: 1038.31 vectors/sec (batch size 1000)
Best read performance: 27.88 queries/sec (1000 queries)Both of these tests have been ran on Apple M1 Pro, similar results have been found on an intel machine.
Image Query
We have an image similarity searching sample
We have measured the performance of Qdrant and Antarys for comparing. This test was inspired from the milvus image similarity test, which can be found here
Here are the results
For Qdrant
For Milvus
For Antarys
Image search results
Qdrant Query Time -> 0.0497 Seconds
Milvus Query Time -> 0.0417 Seconds
Antarys Query Time -> 0.0024 Seconds
Almost a 20x query time bump