Sub-2ms Pre-Filtered
Vector Search at Planet Scale.
In real-world AI & RAG apps, almost every search filters by tenant, permissions, category, or date. When filters narrow the results, OpenSearch 3.9 slows down by up to 19.4× and demands 7× more RAM. MassEngine stays under 2 ms on a fraction of the hardware—so a $99/core/month license saves thousands per month in cloud infrastructure.
Fast When You Filter. 7× Less RAM. A Fraction of OpenSearch’s Cloud Bill.
Standard HNSW indexes (like OpenSearch 3.9) bog down whenever a query filters by customer, date, or tag—and require up to 34.1 GiB of RAM per node just to hold 2M vectors. Evaluated head-to-head on the exact same 2-node × 20-core cluster (40 vCPUs total, RF=2, 8 Shards, 2,000,000 × 1024-d English Wikipedia BAAI/bge-m3 embeddings, 10 concurrent clients):
Everything Modern RAG Needs in One Engine
Replace fragile multi-system pipelines with a single, deterministic search layer.
True Multi-Tenant Isolation
Host thousands of customer workspaces on shared infrastructure with zero noisy-neighbor degradation, scoped RBAC API keys, per-tenant QoS rate limits, and instant GDPR tenant purge.
High-Throughput Reads & Writes
Ingest over 72,000 vectors/sec with instant availability while sustaining up to 6,993 filtered QPS and sub-millisecond query latency under concurrent production load.
Rich Metadata Pre-Filtering
Filter on arbitrary JSON attributes using expressive operators ($eq, $ne, $gte, $lte, $in, $contains), date ranges, and category bitmaps with zero recall penalty.
In-Place Metadata Patching
Update document permissions, pricing, tags, or categories in place via PATCH without re-computing or re-uploading expensive high-dimensional embeddings.
Flexible Dimension Support
Native support for any embedding model dimension—from compact 128-d and 384-d models up to 1536-d and 3072-d frontier embeddings.
1-Click AWS & GCP Clusters
Start free with a 1-Core Developer License (up to 1M vectors), or purchase a signed Ed25519 per-core license (meli_...) and launch a multi-node HA cluster directly inside your own AWS or GCP VPC in 1 click.
Clean REST & SDK Interface
Simple, predictable JSON APIs and typed SDKs for Go, Python, and TypeScript.
POST https://api.massengine.com/v1/search
Authorization: Bearer mk_live_9f8a7b6c...
Content-Type: application/json
{
"vector": [0.021, -0.044, 0.089, "... 1024 dims"],
"k": 10,
"nprobe": 16,
"rerank_pool": 64,
"min_score": 0.75,
"metadata_filter": {
"department": { "$in": ["security", "legal"] },
"clearance_level": { "$gte": 2 },
"tags": { "$contains": "approved" }
}
}
Same 20M Records. 3 Live Hardware Setups.
See how MassEngine serves the exact same 20,000,000 English Wikipedia passages (1024-d embeddings, 83.46 GiB logical corpus across 8 shards) across three distinct live hardware profiles with 100% Top-K recall parity.
2-Node High-End In-Memory
All 8 shards (RF=2) pinned in high-bandwidth RAM across 2 × 20-core nodes (40 cores total, 44.6 GiB RAM/node).
2-Node High-End NVMe
Identical 20M-record dataset on 2 × 20-core nodes using only 2.52 GiB RAM/node via zero-copy mmap NVMe tiering.
4-Node Low-End NVMe
20,000,000 vectors distributed across 4 lightweight Proxmox VMs (1 vCPU / 6.8 GiB RAM each, 4 cores total) backed by NVMe.
Why $99 / Core / Month Saves Thousands vs. OpenSearch
OpenSearch is “free” software that forces you to rent massive, memory-heavy cloud VMs because HNSW indexes consume 7× more RAM and bog down whenever queries include filters. Because MassEngine delivers 15×–32× higher filtered QPS in 7× less RAM (or directly from NVMe SSD), your total monthly cost (Cloud VMs + $99/core license) is up to 85% lower.
OpenSearch 3.9 vs. MassEngine Total Cost of Ownership
| Cluster & Engine | Measured Filtered Throughput | Cluster Hardware & Measured RAM | Est. Cloud VM Cost | MassEngine License | Total Monthly Cost |
|---|---|---|---|---|---|
| OpenSearch (4 Cores Total) | 1M Vectors (RF=2) • 21–36 QPS (5%–10% filter) | 4 × 1 vCPU VMs • 6.8 GiB RAM/node | ~$140 / mo | $0 | $140 / mo (16×–23× slower QPS) |
| MassEngine (4 Cores Total) | 1M Vectors (RF=2) • 500–594 QPS (2.6–3.1 ms) | 4 × 1 vCPU VMs • 2.33 GiB RAM/node | ~$140 / mo | $396 / mo (4 cores) | $536 / mo Outperforms 40-Core OpenSearch 3.9 |
| OpenSearch 3.9 (40 Cores Total) | 2M Vectors (RF=2) • 216–392 QPS (18.5–22.3 ms) | 2 × 20-Core Nodes • 34.1 GiB RAM/node | ~$1,850 / mo | $0 | $1,850 / mo (Still slower than 4-core MassEngine) |
| MassEngine (40 Cores Total) | 2M Vectors (RF=2) • 5,849–6,993 QPS (1.15–1.52 ms) | 2 × 20-Core Nodes • 4.88 GiB RAM/node | ~$960 / mo (7× less RAM) | $3,960 / mo (40 cores) | $4,920 / mo vs. $25k+/mo for 15× OpenSearch nodes |
| OpenSearch 3.9 (20M Corpus) | 20M Vectors (1024-d) • High-RAM HNSW | Requires ~340 GiB RAM per copy (~680 GiB RF=2) | ~$4,500+ / mo | $0 | $4,500+ / mo |
| MassEngine 20M NVMe (4 Cores Total) | 20M Vectors (1024-d) • Live Setup 3 (--nvme-mode) |
4 × 1 vCPU VMs • 1.27 GiB RAM/node | ~$140 / mo | $396 / mo (4 cores) | $536 / mo Save ~88% on 20M Vectors |
1-Core Developer — $0 / mo
Instant activation in the Cloud Console. Run on your laptop, Docker, or any cloud VM for development, evaluation, and workloads up to 1,000,000 vectors.
Per-Core License — $99 / core / mo
Pick your number of nodes and cores per node. Scale out or scale up anytime across your floating core pool with no memory or storage surcharges.
Need enterprise volume licensing, custom invoicing, or architecture guidance? Reach out via the Enterprise form below ↓
Contact Enterprise Engineering
Planning an In-VPC deployment or 100M+ vector cluster? Reach out below—or create a self-service Cloud Console account to start immediately.