Are you an LLM? Read llms.txt for a summary of the docs, or llms-full.txt for the full context.
Skip to content
Get started

Concepts

Mubit groups similar calls into tasks and uses kept examples to train and test adapters.

Not every call is kept, and not every version passes. Training also depends on the task's limits and eligibility.

The diagram in words

Mubit groups recorded calls into tasks and selects kept examples. Once enough eligible examples are available and training can run, Mubit trains and tests an adapter version. A version then goes through a rollout, after which it can answer some of the task's calls in place of your provider. If a version is below target, review its results and provide more useful examples.

Four terms to know

  • Request: One call your app sends through Mubit. Recorded calls appear in Requests.
  • Task: A recurring kind of request. Calls with matching model, instructions, tools, tool choice, and output format form a task.
  • Kept example: A call Mubit accepts for training or testing. Not every recorded call becomes a kept example.
  • Adapter: A small model trained for one task. Each training creates a version, which Mubit tests against the provider model's answers.

You can compare a version with your provider model in Try models. After its tests, a version goes through a rollout: a silent test, a partial rollout, then live. From then on Mubit can answer some of the task's calls with the adapter, under the model ID your app sends. Passing a test does not guarantee suitability for every input; review representative examples from your application. Train adapters explains each step.

Two different keys

Your app uses a Mubit key to call Mubit. Mubit uses a provider API key to call the provider: the key of your own provider account. Where your workspace offers it, a Neon AI Gateway key sends the call to your Neon branch's AI Gateway instead.

A project holds your calls, tasks, adapters, and keys. Project administration is separate from this learning flow.

Runtime calibration

An app can store what it has learned and recall what fits on later runs, and report whether a run worked. It uses the same Mubit key and host as your model calls. Runtime calibration shows how.

Next: Create your first adapter.

For detail: Tasks and sessions · Mubit keys · Provider keys.