Search your documents by meaning, not keywords.
Korvo Embed understands your documents on your machine. Nothing is uploaded — not your files, not your searches, not a single embedding.
34 MB
model, downloaded once
384
dimensions per chunk
512
tokens per chunk
0
bytes of your text sent
What it does
Retrieval that reads for intent.
Meaning, not keywords
Ask "how do I change my password" and find the page that says "reset your credentials." No shared words needed.
Entirely on your machine
A 12-layer transformer quantized to 34 MB, running on your CPU. No server, no account, no rate limit, no cost per query.
Downloads once
One 34 MB model on first index, fetched by cryptographic hash and verified before use. Then the network is never touched.
Private by architecture
A privacy policy is a promise. This is a property of the binary.
The component that reads your documents is structurally incapable of sending them anywhere — and it is checked on every build.
$ cargo tree -p korvo-embed --prefix none | grep -E 'reqwest|hyper|ureq|curl'
(no output)Never compiled in
The exfiltration capability does not exist in the binary. Model download lives in a separate component that never sees a document.
Content-addressed model
Fetched by cryptographic hash, not by URL. A compromised CDN cannot substitute a different model.
Your index is yours
Vectors are files on your disk. No proprietary format behind a service, nothing to lose if we disappear.
Availability
Shipping today. Licensable now. API next.
Shipping inside the Korvo macOS app. Every project you index is searched by meaning, with no configuration.
Licensed for embedding into your own product, with integration and evaluation work for local-first and compliance-bound teams.
Korvo API Platform — the same engine and the same provenance guarantees, as a managed building block.
Roadmap
Faster first index — multi-core inference and batching, with no re-indexing required.
Incremental indexing — only re-embed what changed, so the full pass happens once.
Reranking — a second stage to order near-tied top results. The largest quality headroom we have measured.
Cloud sync of embeddings · non-English models · Intel support · training on user data.
For security review
Built to survive technical due diligence.
- Data residency
- Documents never leave the endpoint. The indexing and search component contains no network client.
- No third-party processing
- No model provider, no inference API, no telemetry on document content. Nothing to add to a subprocessor list.
- Auditable supply chain
- The model is pinned to an upstream revision and verified by SHA-256. Provenance ships in a machine-readable manifest.
- Reproducible
- The shipped model can be rebuilt from the pinned revision and checked to be byte-identical.
Scope
What we are careful not to claim.
An engine sold on a privacy guarantee cannot afford one overstatement.
Apple Silicon
macOS on Apple Silicon. No Intel build.
English
Other languages produce vectors, and they will be poor.
512 tokens per chunk
Roughly two pages. Longer input is truncated. No cap on corpus size.
Text in
Extracting text from PDFs, images or audio is a separate step.
One download
Not "fully offline" — 34 MB arrives on first index, then never again.
Engine-scoped
The no-network guarantee covers search and indexing, not the whole Korvo app.
Partnership
Want this inside your product?
We license Korvo Embed to teams building local-first software, and to anyone whose security review will not permit documents to reach a third-party inference API. Tell us what you are building and we will tell you honestly whether it fits.
Use it today
Korvo Embed powers semantic search in the Korvo macOS app. No setup, no API key.
Download Korvo →Embed it in yours
- Engine licensing for on-device retrieval
- Integration and evaluation on your own corpus
- Provenance and model-card material for security review
- Design-partner access to the Korvo API Platform
Questions
The ones people actually ask.
How do I know it isn’t uploading my files?
Check it yourself. A dependency audit of the engine returns no HTTP client, and that check runs as a build gate on every commit. It is a property of the binary, not a promise.
Isn’t local search worse than cloud AI?
For retrieval, a 33M-parameter encoder runs fine on a laptop. You give up a model too big to run locally; you get documents that never leave the machine, no per-query cost, and search that works on a plane.
Why a 34 MB download if it’s local?
That is the model — the thing that understands language. Downloaded once, verified by cryptographic hash, then never contacted again.
What happens offline?
Everything, once the model is downloaded. Before that, keyword search works and semantic search says so plainly.
On-device embeddings via bge-small-en-v1.5, int8-quantized, 384 dimensions. Apple Silicon.