Filtered vector search under 2 ms—in RAM or straight from NVMe.
Graph indexes slow down when queries filter by tenant, permissions, or metadata, and require expensive RAM to stay responsive. MassEngine runs inside your own VPC using 7× less memory than HNSW, with zero multi-hour compaction stalls.
Built for production filters and predictable cloud bills.
Measured on AWS EC2 with unmodified upstream VectorDBBench (Cohere-10M) and multi-node Wikipedia datasets (2M to 144M vectors).
| Workload | Typical HNSW Baseline | MassEngine |
|---|---|---|
| Filtered search (10% match, 2M vectors, 40 vCPUs) | 392 QPS • 18.5 ms (OpenSearch 3.9) | 5,849 QPS • 1.5 ms (4.9 GiB vs. 34.1 GiB RAM/node) |
| 10M index build time (Cohere-10M, 16 vCPUs) | 2.4 – 4.0 hours (Milvus / Qdrant 16c64g) | 11.1 minutes on c6id.4xlarge (32 GiB RAM, NVMe) |
| Streaming reads under 1,000 writes/sec | 156 – 348 QPS (Milvus / Qdrant 16c64g) | 1,797 QPS • 2.4 ms p99 (0.985 recall) |
| 20M-vector cluster monthly cost (1024-d, AWS) | $4,500+ / mo (RAM-bound HNSW instances) | $536 / mo all-in (4 NVMe VMs + 4-core license) |
Drop into your stack in five minutes.
Single static binary for Linux and macOS with zero external dependencies, paired with official zero-dependency SDKs.
pip install massengine
npm install @massengine/client
go get github.com/massengine-dev/massengine-go@latest
from massengine import MassEngineClient, Doc
client = MassEngineClient(
base_url="http://localhost:8080",
api_key="mk_live_9f8a7b6c...",
)
client.ingest("acme-prod", [
Doc(id="doc-1", vector=[0.021, -0.044, 0.089], metadata={"department": "security", "clearance": 3})
])
res = client.search(
"acme-prod",
vector=[0.021, -0.044, 0.089],
k=10,
metadata_filter={"department": {"$in": ["security", "legal"]}, "clearance": {"$gte": 2}},
)
Filter-first execution
Metadata predicates ($eq, $in, $gte, $lte, $contains) narrow candidates before vector scoring—queries accelerate as filters tighten, and attributes can be patched in place without re-embedding.
Multi-tenant isolation
Isolate thousands of customer workspaces on shared hardware with scoped RBAC keys, per-tenant rate limits, and instant tenant deletion without background segment merges.
Runs inside your VPC
Your embeddings and documents never leave your network. Deploy on AWS, GCP, Azure, Docker, or bare metal with zero-downtime rolling upgrades.
Simple per-core licensing. Zero RAM or storage fees.
Bring your own cloud or bare metal. Stateless coordinator gateways consume zero licensed cores.
Free forever for local development, prototypes, and single-core workloads.
- 1 CPU core (single node)
- Up to 1,000,000 vectors
- RAM and Direct NVMe modes included
- No credit card required
Floating core pool across any multi-node HA topology in your own VPC.
- Choose any mix of nodes and cores
- Unlimited vectors, RAM, and NVMe storage
- 1-click AWS CloudFormation & GCP Terraform
- Zero-downtime rolling cluster upgrades