CEREBRAL · The vector + graph database for AI search

How CEREBRAL works

One query. One engine. Five things happening at once.

01 · Ingest

Documents become embeddings, compressed from the first write.

Chunks from your contracts, manuals and tickets are embedded and compressed the moment they're stored. Nothing waits for a batch job. Search is built to run on the compressed data directly, designed to be served from disk.

Compressed on write
Built to search while compressed

write path

MSA-2291.pdf

§14.3 Late delivery. Where the Supplier fails to deliver…

X200 service manual

4.2 Fault E-4021: the pump stops shortly after start-up…

Ticket · pump stops

Site reports the X200 halts after the 3.2 firmware…

embedcompress

Index · harbour-ops

MSA-2291 · §14.3
X200 manual · §4.2
Ticket · pump stops

stored size

full-size vectors (reference)

stored by CEREBRAL

compressed on write · built to search while compressed · designed to be served from disk

02 · Meaning · illustrative

Matched by meaning.

Questions are matched by meaning, even when the words differ. (Illustrative example.)

Meaning catches the phrasing you didn't use

querypump fault E-4021 after firmware 3.2

Meaning

pump stops after update

halts after flashing firmware

start-up fault alarm

rank

Results · ranked · example data

X200 service manual · §4.2

…fault E-4021: pump stops after firmware 3.2 update…

meaning

Ticket · pump stops

…halts after flashing the new firmware on site…

meaning

Release notes · 3.2

…firmware 3.2 changes pump start-up timing…

03 · Graph hops

Find it by meaning, then follow the relationships to the answer.

Embeddings sit on graph nodes. One query finds the relevant passage, then walks its links between people, products, contracts and events. That's GraphRAG without a second database to keep in sync.

Entry point found by meaning
Hops follow real relationships
The path is the explanation

askWhich contracts are exposed if Arden misses the March delivery?

Delay notice · 2 Mar

Arden Castings advises the March shipment will slip…

found by meaning

from

supplies

covered by

cites

owns

eventDelivery delay · Mar

supplierArden Castings

productPump X200

contractMSA-2291

contractMSA-2307

clause§14.3 Late delivery

personPriya Shah

productValve V10

contractMSA-1180

Exposed: MSA-2291 and MSA-2307. Both cover Pump X200 and cite §14.3 late delivery. Account lead: Priya Shah.example · one query, graph hops

04 · One engine, one API

Retire the second database and the glue between them.

A stitched stack runs a vector store, a graph database and a job to keep them in step. CEREBRAL keeps vectors and relationships in one backend, so one request can match, rank and follow links.

One system to run and secure
No sync job to babysit
One request for match, rank and hops

Stitched stackCEREBRAL

Your app or agent

glue code

Vector databaseembeddings

Graph databaserelationships

⇄ sync job

vectorsrelationships

one API

a job to keep them in step

two systems to secure and pay for

one request

one system to secure

// illustrative only · API shape not final
const res = await crbrl.query({ text: "pump fault E-4021", match: ["meaning"], follow: { hops: 2 },
});

05 · Deploy

Run it where your data has to live.

Use managed cloud (roadmap), or deploy on your own servers. Built for on-prem and sovereign deployment, including in Australia. Same engine, same API, same queries. Sovereign and on-prem deployments come with an enterprise licence.

Regulated data stays inside your boundary
No lock-in to one region or vendor
Pilot first, then roll out

stays in your boundary

Same engine, wherever it runs

Same API, same queries
Vectors and graph together
Compressed from the first write
Per-tenant projects for clients and teams

Sovereign and on-prem: enterprise licence. Pilots by arrangement.