⚡ World’s Fastest Vector Database for AI & RAG

Casper

Casper is a high-performance Vector Search Database, perfectly suited for high-load search systems and AI applications (RAG). It provides a robust and scalable solution to store, search, and manage vectors efficiently.

Casper is built using Rust 🦀 for performance and reliability. Casper clients Python • Go • Rust

---

Why Casper ?

Casper is the fastest vector database in our internal benchmarks. It consistently outperforms Qdrant across Top@K workloads and both f32 and i8 quantizations. Notably, Qdrant is widely recognized as the leading open‑source engine and demonstrates state‑of‑the‑art throughput versus other databases (e.g., Weaviate, Milvus), as shown in their published results: Qdrant benchmarks. Surpassing Qdrant therefore places Casper ahead of the current open‑source performance leader.

In practice, Casper delivers up to an order‑of‑magnitude higher RPS compared to Qdrant on our datasets, which translates directly into substantial infrastructure savings: fewer CPU cores and instances to achieve the same SLA, lower memory pressure, and reduced total cost of ownership due to more efficient use of compute resources. Casper is the ideal solution for high-load systems, real-time search, and AI & RAG.

Conclusion: Casper achieves performance unattainable for other databases under comparable conditions, requires fewer compute resources at the same load, and materially reduces infrastructure costs through more efficient CPU and memory utilization.

Casper vs Qdrant

Benchmarks RPS & Recall

Hardware:

• CPU: AMD EPYC 9474F (48 cores 96 threads L3 Cache 256 MB)
• Memory: 256 GB RAM

Dataset: deep-image-96-angular.hdf5

• Vectors: 9,990,000
• Dimension: 96
• Metric: Inner Product (vectors are L2-normalized)

HNSW

• m: 16
• ef construct: 200

Qdrant configured with quantile 0.99 (for int8), always ram enabled.

Search-time parameter. For every measurement in the tables below we explicitly set ef_search = limit on both engines (for Qdrant via params.hnsw_ef, overriding its server-side default of max(limit, 128)). This is the smallest valid HNSW ef and the most apples-to-apples comparison: both engines do the minimum amount of graph exploration the algorithm allows.

Index granularity (why Qdrant recall is higher). Casper builds a single monolithic HNSW index per collection. Qdrant splits the collection across multiple segments (segments_count: 8 in this benchmark) and runs HNSW search independently in every segment, then merges the per-segment top‑K on the coordinator. With per-segment hnsw_ef = limit, Qdrant effectively examines segments_count × limit = 8 × limit candidates per query — eight times more than Casper for the same nominal ef. The higher recall Qdrant shows at every K in the tables therefore reflects this storage organization, not better HNSW graph quality; the same effect is the reason its RPS is correspondingly lower (more work per request).

Full Precision (F32)

Requests per second, RPS

Engine	Top@10	Top@100	Top@1k	Top@10k	Top@100k
Casper	248.86 k	93.55 k	12.02 k	1.43 k	163
Qdrant	13.25 k	15.65 k	2.48 k	285	28
Speedup	18.8x	6.0x	4.8x	5.0x	5.8x

Recall

Engine	Top@10	Top@100	Top@1k	Top@10k	Top@100k
Casper	0.606	0.854	0.957	0.987	0.995
Qdrant	0.762	0.953	0.993	0.999	1.000

Scalar Quantization (I8)

Requests per second, RPS

Engine	Top@10	Top@100	Top@1k	Top@10k	Top@100k
Casper	248.79 k	145.53 k	22.47 k	2.36 k	221
Qdrant	26.90 k	19.43 k	3.29 k	345	28
Speedup	9.2x	7.5x	6.8x	6.8x	7.9x

Recall

Engine	Top@10	Top@100	Top@1k	Top@10k	Top@100k
Casper	0.578	0.820	0.921	0.957	0.975
Qdrant	0.729	0.914	0.959	0.975	0.985

HNSW

Casper features a highly efficient HNSW (Hierarchical Navigable Small World) index, providing fast and accurate similarity search.

Metrics

Casper supports multiple distance metrics:

• Euclidean
• L2SQ
• Cosine
• Inner-Product

Quantizations

Quantizations: f32 (full precision), i8 scalar quantization — reducing memory footprint and improving search performance.

• F32
• I8

---

Free Access

Casper is currently completely free. You can use the following free API token to run Casper:

export API_TOKEN=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJleHAiOjE3OTMyOTAzNTMsImZyZWUiOnRydWV9.GxqiVw5kPzmPb25vo2CMOEwnBhjTH_GTAHeDg_nhlIQ

Quick Start

Download and Launch

To quickly get started with Casper, follow these steps:

1. Download the latest release:

wget https://github.com/casper-vdb/casper/releases/download/v0.1.0/casper-x86_64-unknown-linux-gnu.tar.gz

2. Extract the downloaded archive:

tar -xzvf casper-x86_64-unknown-linux-gnu.tar.gz

3. Set API token:

export API_TOKEN=<YOUR_API_TOKEN>

4. Run Casper:

./casper

Now you're ready to use Casper and explore its features!

Docker: Download and Launch

1. Pull the image:

docker pull alexryzhickov/casper:latest

2. Set API token:

export API_TOKEN=<YOUR_API_TOKEN>

3. Run the container:

docker run -d --name casper -p 7222:7222 -p 7223:7223 -e API_TOKEN="$API_TOKEN" alexryzhickov/casper:latest

4. Verify health:

curl http://localhost:7222/health

Clients

Casper provides client libraries for several programming languages:

• Python
• Go
• Rust

API Documentation

Casper exposes an HTTP & GRPC API for managing collections, indexing (HNSW), inserts/updates/deletes, and search. For full endpoint descriptions and curl examples, see the documentation:

• API Docs
• Configuration Docs

Cluster Mode

Casper supports a master/replica cluster topology with WAL-based replication. The master accepts writes and broadcasts them to connected replicas over gRPC; replicas replay the stream and serve reads.

By default the server runs in standalone mode. To enable clustering, pass --cluster-role and configure replication. The full set of flags is documented in docs/config.md.

Run standalone

A single-node deployment with no replication — the default mode:

./casper \
  --host 127.0.0.1 \
  --http-port 7222 \
  --storage-dir ./storage

Run a master

The master accepts writes on --http-port and listens for replica connections on --grpc-port. With --sync-replicas N it waits for N replica fsync acks before returning 200 from the endpoints that perform upsert and delete operations:

./casper \
  --host 127.0.0.1 \
  --http-port 7222 \
  --grpc-port 7223 \
  --cluster-role master \
  --storage-dir ./storage-master \
  --sync-replicas 1 \
  --write-ack-timeout-ms 5000

If fewer than --sync-replicas are connected, writes fast-fail with 504 InsufficientReplicas. If the deadline elapses, they fail with 504 ReplicationTimeout.

Run a replica

A replica dials the master's --grpc-port via --master-addr:

./casper \
  --host 127.0.0.1 \
  --http-port 7322 \
  --storage-dir ./storage-replica \
  --cluster-role replica \
  --master-addr 127.0.0.1:7223

The replica's on-disk layout mirrors the master's, so a replica can be restarted with --cluster-role=master for promotion.

---

Features

• Advanced Vector Search: High-speed retrieval for complex AI-driven applications.
• Scalability: Designed to handle large-scale data with ease.
• Robust and Reliable: Built in Rust for high performance even under heavy loads.

---