completers

You can streamline text generation for RL rollouts by using a TokenCompleter interface, which wraps a SamplingClient to return raw tokens and optional logprobs for precise control during training and evaluation.

Token-level interfaces return raw tokens and logprobs for low-level RL training control, while message-level interfaces accept and return structured Message objects for high-level multi-turn conversations and evaluation workflows.

Yes, you need a SamplingClient to generate structured messages. Both the MessageCompleter and TokenCompleter abstractions require a SamplingClient instance to manage rendering, stop conditions, and error handling during text generation.

Yes, you can use completers for both RL evaluation and conversational tooling. They provide flexible interfaces supporting both precise token access for RL training and structured message handling for multi-turn conversations.

To get logprobs during RL training rollouts, use the TokenCompleter interface which returns optional logprobs alongside generated tokens, enabling precise probability tracking for policy gradient updates and evaluation.