iRaven // Huginn
Huginn
Access to models for the people of a tenant's app, with no provider's key inside the app.
What it is
Huginn gives the people of a tenant's app access to models, with no provider's key inside the app. A provider is a service that runs models, and a call is one request from an app to a model, through iRaven. The app holds no secret: at most a publishable key, which only names the tenant and does not need to be hidden. A tenant's server can call too, with a secret key of its own, and it names the person it acts for. The name is from Norse myth, where Huginn is one of the two ravens of Odin, and it means thought. The other raven, Muninn, names iRaven's memory for a person's AI assistant.
How it works
-
The app asks for a name
A tenant gives models names of its own, such as
voiceorplanning. The app asks for a name, not for a model. Each name points at a model and never at a provider, so the tenant can change the model behind a name without a new version of the app. -
Which calls iRaven accepts
Chat calls use the common formats that existing libraries already send, so a library only needs iRaven's address. Other calls, such as speech, pictures and video, must be on a fixed list, and iRaven refuses any call that is not. Long work, such as a video, runs as a job: the app gets the job's number and fetches the result when it is ready.
-
Where the models come from
iRaven provides models through providers that it sets up, and a tenant sees the model and its name, never the provider. A tenant's own providers, shared by the people of its app, are planned and not designed yet.
-
What iRaven keeps
iRaven keeps the counts and sizes of each call. It never stores or logs what a person asked or what a model answered.
-
A tool that serves it
Sign-in and devices
Each tenant has its own sign-in. A person signs in with a passkey, a password, a code sent by email, or an account at an outside sign-in service. A token is a short-lived pass that names the person. iRaven ties it to a key that the person's device makes. Every call carries a fresh proof signed by the device, so a copied token is of no use on another device.
-
A tool that serves it
Usage and ceilings
A ceiling is the most that may be used in a period. A tenant sets ceilings for a person, a group, a key, a name or the whole tenant, per day, week or month, or over a moving window of hours or days. Before a call runs, iRaven holds back an estimate against every ceiling that applies, and when the call ends, iRaven settles the real amount. A hard ceiling refuses any call that would go past it.
What is built and what is not
- Models that iRaven providesBuilt, not yet deployed
- A tenant's own providers, shared by the people of its appPlanned
The models that iRaven provides are built and not yet deployed. A tenant's own providers are planned and not designed yet. Its two tools are built and not yet deployed: sign-in and devices, and usage and ceilings.