Status
Live
Live · trained & serving
AGPT-1
AGPT-1 is our first model, and the proof of the whole thesis: there is no foundation model underneath it. We trained it from random weights on our own data, and the result — tokenizer, architecture, and weights — is owned outright, with no base-weight license to answer to.
What it is
A from-scratch text-embedding model.
AGPT-1 turns text into vectors — the foundation for the semantic search we're bringing to the platform. It was pre-trained by us on a license-clean corpus, with weights initialized from random, not fine-tuned from someone else's base.
That distinction is the whole point: every other shortcut starts from a licensed foundation model. AGPT-1 starts from nothing but our data and our design, so the resulting IP is unambiguously ours.
Type
Text embeddingstretch-embed
Architecture
CustomBuilt in-house
Tokenizer
ProprietaryOur own
Training
In-houseOn our own data
Base model
NoneRandomly initialized
Ownership
100%No base-weight license
Why from scratch
Owned weights change what we can do.
01
Clean IP
No upstream license terms, usage caps, or model-provider dependency. The weights are an asset we own.
02
Our tokenizer
A proprietary vocabulary trained on our own corpus, tuned to our domain language rather than a generic web crawl.
03
Efficient by design
Compact enough to train and serve efficiently, and fast enough to embed at the scale retrieval will need.
04
The spine
The same from-scratch pipeline extends to our other models — AGPT-1 is the template, not a one-off.