completers

What problem does it solve?

Completers provide convenient, structured interfaces for text generation by wrapping a SamplingClient, enabling both low-level token access and high-level message handling during RL rollouts and evaluation.

Core Features & Use Cases

• TokenCompleter: low-level interface that returns tokens and optional logprobs for precise control during RL training.
• MessageCompleter: high-level interface that accepts and returns Message objects for multi-turn conversations and evaluation.
• Flexible usage: supports integration in RL rollouts, tool-using environments, and evaluation workflows; ideal when you need either raw tokens or structured messages.

Quick Start

Instantiate a TokenCompleter or MessageCompleter with a SamplingClient and run a simple prompt to generate either tokens or a parsed Message.