Extend Your Database with Premium Plugins
GPU acceleration, in-database machine learning, and adaptive caching. Add or remove at any time - no downtime, no migration.
GPU Acceleration
CUDA-based GPU acceleration for HNSW vector similarity search. Batch distance computation, brute-force KNN, graph construction, SQ8 quantized distance, priority scheduling, multi-query batching, and exact re-ranking - all in-process on database nodes with zero network hops.
Supported Operations
Batch Distance
Compute distances between query vector and candidate set
10-50x speedup
Brute-Force KNN
Exact nearest neighbor search over full dataset
50-100x speedup
Graph Construction
Build HNSW graph layers in parallel
5-20x speedup
SQ8 Quantized
Distance computation on scalar-quantized vectors
20-80x speedup
Priority Scheduling
Multi-stream CUDA scheduler with Critical, Normal, and Background priority lanes and anti-starvation promotion
3 priority levels
Multi-Query Batching
Coalesces up to 64 concurrent search queries into a single GPU kernel launch, amortizing transfer overhead
64 queries/batch
GPU Re-Ranking
Exact float32 re-ranking of approximate search candidates with rank delta tracking for quality metrics
10x candidate pool
Tier Comparison
| Feature | Standard $799/mo | Pro $1499/mo | Enterprise $3999/mo |
|---|---|---|---|
| Batch Distance | |||
| Brute-Force KNN | |||
| Multi-Query Batching | |||
| Graph Construction | - | ||
| SQ8 Quantized | - | ||
| Priority Scheduling | - | - | |
| GPU Re-Ranking | - | - | |
| Infrastructure | |||
| GPU Memory (VRAM) | 2 GB | 4 GB | 8 GB |
| Max Concurrent Queries | 32 | 128 | 512 |
Feature Deep-Dives
ML Engine
Run machine learning directly inside your database — 13 algorithms including regression, clustering, classification, dimensionality reduction, anomaly detection, and ensemble methods. Pure Rust, zero external services.
13 Built-in Algorithms
Example Workflow
// Train a KMeans model on user behavior
db.ml.train({
algorithm: "KMeans",
source: "user_events",
features: ["session_duration",
"pages_viewed", "actions_count"],
params: { k: 5 }
})
// Predict cluster for new data
db.ml.predict({
model: "user_events_kmeans",
input: {
session_duration: 120,
pages_viewed: 8,
actions_count: 15
}
})Tier Comparison
| Feature | Standard $99/mo | Pro $199/mo | Enterprise $299/mo |
|---|---|---|---|
| Algorithms | All 13 | All 13 | All 13 |
| Max Training Samples | 1M | 10M | Unlimited |
| Max Features | 5,000 | 10,000 | Unlimited |
| Max Iterations | 1,000 | 1,000 | 10,000 |
| Model Persistence | |||
| Batch Prediction API | -- | ||
| Custom Iteration Limits | -- | -- |
Intelligent Cache
Production-grade 3-tier adaptive caching with ML-inspired eviction scoring (7 signals), automatic workload classification (5 classes), and predictive pre-warming. L1 memory, L2 SSD, L3 distributed — sub-millisecond latency, zero config.
L1 In-Memory
moka hash map
<1ms
latency
L2 SSD
LZ4 compressed
1-5ms
latency
L3 Distributed
TCP mesh
5-20ms
latency
ML-Inspired Eviction Scoring - 7 Signals
A production-grade eviction system that replaces simple LRU/LFU with a 7-signal composite scoring function. Higher scores mean the entry stays longer.
Recency20%
Exponential decay based on time since last access
Frequency25%
Log-scaled access count normalized across entry ages
Cost20%
Recompute cost - cold tier queries score higher than L1 hits
Predicted Value25%
Future access probability from the Access Pattern Predictor
Size Efficiency10%
Inverse size - smaller entries are more space-efficient
Age Decay0%
Optional - prevents zombie entries cached once and never evicted
Settling Boost2x
New entries get a score boost during settling period (30s)
Cache Types
Documents
Individual document lookups by ID
Cold Partitions
Frequently accessed cold-tier data promoted to cache
Query Results
Cached results for repeated aggregate/filter queries
Aggregations
Pre-computed aggregation pipeline results
Metadata
Collection schemas, index definitions, and config - tiny but accessed on every query
Tier Comparison
| Feature | Standard $299/mo | Pro $599/mo | Enterprise $999/mo |
|---|---|---|---|
| L1 In-Memory Cache | |||
| L2 SSD Cache | |||
| L3 Distributed Cache | -- | -- | |
| Eviction Strategy | LRU + Frequency | ML Scoring (7 signals) | ML Scoring (7 signals) |
| Workload Classification | -- | 5 classes | 5 classes |
| Predictive Pre-Warming | -- | -- | |
| Admission Control | Size-based | Size + Value | Size + Value + Workload |
| Cache Types | Docs + Query | All 5 types | All 5 types |
| Cross-Node Coordination | -- | -- | |
| Infrastructure | |||
| L1 Memory (max entries) | 100K | 500K | 1M |
| L2 SSD Storage | 10 GB | 50 GB | 200 GB |
| L3 Distributed | -- | -- | Unlimited |
Feature Deep-Dives
Plugin Pricing at a Glance
Mix and match plugins across your clusters. Add or remove at any time with no downtime.
| Plugin | Tier 1 | Tier 2 | Tier 3 |
|---|---|---|---|
| GPU Acceleration | $799/mo Standard | $1499/mo Pro | $3999/mo Enterprise |
| ML Engine | $99/mo Standard | $199/mo Pro | $299/mo Enterprise |
| Intelligent Cache | $299/mo Standard | $599/mo Pro | $999/mo Enterprise |
Build AI Applications on a Database That Gets It
One database for your vectors, your context, your agent memory, and your application data. Start free. Scale without limits.