The vector DB you can actually trust.

Full-fidelity recall at any scale. A native on-disk graph that scales without giving up accuracy — and the same REST shape your team already uses, so migration is a URL swap, not a rewrite.

A vector database you can actually trust.

Full-fidelity recall, on disk, at any scale — no shortcuts, no accuracy trade-offs. A drop-in replacement for your existing vector store, with the same REST shape you already know.

Full-fidelity recall, on disk

No shortcuts, no trade-offs. A native on-disk graph that scales without giving up accuracy. Cosine, and Euclidean metrics supported.

POST /v1/vector/indexes/my-idx/query
{
  "vector": [0.12, 0.43, ...],
  "top_k": 10,
  "metric": "cosine"
}

Index management

Familiar index primitives: create, list, describe, delete. Pick your dimension and metric at creation; per-index quotas enforced at the engine level.

POST /v1/vector/indexes
{ "name": "products", "dimension": 1536, "metric": "cosine" }

Bulk CRUD + paginated scrolls

Upsert thousands of vectors per call. Query, fetch by ID, delete in bulk, and scroll through entire indexes — everything you need to run a production vector store.

POST /v1/vector/indexes/my-idx/vectors
{
  "vectors": [
    { "id": "doc-1", "values": [...], "metadata": { "category": "ai" } },
    { "id": "doc-2", "values": [...], "metadata": { "category": "db" } }
  ]
}

Payload / metadata filtering

Store arbitrary JSON payload per vector; filter at query time with operators like $eq, $lt, $gt, $in. Filtered queries stay fast — filters fuse with the search itself.

{
  "vector": [...],
  "top_k": 10,
  "filter": {
    "category": { "$eq": "electronics" },
    "price": { "$lt": 500 }
  }
}

Tenant isolation + quotas

Each tenant's indexes are namespaced with composite-ID encoding at the engine level — no cross-tenant leakage, even under collision. Hard tier caps, lock-free counters.

REST and gRPC, side by side

Reach for REST when you want quick integration, or native gRPC for throughput-sensitive workloads. The HTTP shape is a drop-in replacement for your existing vector store — most clients work with minimal changes.

100% recall. No shortcuts, no trade-offs.

Most vector databases quietly trade accuracy for speed. Ours doesn't — the graph is built once, lives on disk, and serves full-fidelity results from the first query onward.

Concern
How the index lives
Most vector DBs
Loaded into memory, or worse
keyes.ai
Native on-disk graph — no memory hacks
Concern
Recall for agent memory
Most vector DBs
80–95% (approximate)
keyes.ai
100% — every vector, every time
Concern
Behavior at scale
Most vector DBs
Falls back to brute force as data grows
keyes.ai
Stays indexed — no slow path, ever
Concern
Cold-start latency
Most vector DBs
Seconds (load from disk)
keyes.ai
Sub-millisecond — ready instantly
Concern
Accuracy with filters
Most vector DBs
Degrades with filter + re-rank
keyes.ai
Constant — filters and search stay fused

Drop it in where you already use a vector store.

Same familiar API shape, better recall, smaller bill. Migration is a URL swap — not a rewrite.

RAG pipelines

Embed your docs, store the vectors, retrieve the right ones. 100% recall means the relevant chunk is always in the top-k — no "why did the LLM miss that?" moments.

Recommendation systems

Content-based or collaborative filtering with millisecond latency. Payload filters let you scope by availability, price, or cohort without a second query.

Similarity search (products, images, docs)

"Find me items like this one." Drop in your embeddings, pick your metric (cosine / dot / Euclidean), and get scored top-k results with metadata attached.

Duplicate & near-duplicate detection

Find the paraphrased post, the almost-identical image, the copy-pasted commit message. 100% recall keeps the near-misses from slipping through.

Pay for vectors stored, not throughput.

Query as fast as you need — your bill is sized by what you keep, not what you compute.

Free

For prototypes & learning

  • 1 index
  • 10K vectors
  • 1536-dim cosine
  • Full recall, on-disk graph

Starter

For side projects & early RAG

  • 3 indexes
  • 100K vectors
  • 1536-dim cosine
  • Payload filters included

Enterprise

For regulated orgs, self-hosted

  • Unlimited
  • Unlimited
  • 1536-dim cosine
  • Self-hosted in your cloud

Ready to build with
keyes.ai?

Join the private beta. Get early access to GitDB, Memory, Vector, and Embedded Robotics services.