Plural
Sign in

Documentation

Start with one routed request. End with a repeatable evaluation loop and scored traces you can use to improve an agent.

Plural connects production routing, evaluation, and training data around one trace format. Follow the path below in order if you are starting fresh. Jump to the outcome you need if you already have traffic running.

  1. 01Send your first requestInstall the client, connect a workspace key, and choose hosted or bring-your-own provider keys.
  2. 02Browse the available modelsSee model ids, context windows, modalities, parameters, providers, regions, and prices.
  3. 03Swap models without rewriting the applicationKeep one request and response shape across providers, then add fallbacks when you need them.
  4. 04See what happened and keep the useful recordsInspect production calls and eval episodes, control sensitive content, label outcomes, and create datasets.
  5. 05Turn your task into an evaluation harnessDefine state, observations, tools, stopping, and reward once so any model or policy can attempt the same job.
  6. 06Choose a winner and catch regressionsRun the same tasks across models or policies, compare scored reports, and retain episodes for later RL.

Running in production? Keep error handling beside the routing guide.