Adapters

An adapter sits between a module and the language model. It turns a signature plus inputs into prompt messages, and parses the model's reply back into the signature's domain output. You rarely call an adapter directly; you choose which one is active.

dspy4s ships two:

• ChatAdapter formats the task as text and parses a text reply. It is the default and works with any model.
• JSONAdapter asks the model for structured JSON output, which is more reliable for rich output types when the model supports it.

Choosing an adapter

The adapter is taken from the runtime context. These two functions run the same structured-output program under each adapter:

def newsWithChatAdapter(using RuntimeContext): Either[DspyError, List[ScienceNews]] =
  withAdapter(ChatAdapter())(runNews)

def newsWithJsonAdapter(using RuntimeContext): Either[DspyError, List[ScienceNews]] =
  withAdapter(JSONAdapter())(runNews)

The program and signature are identical. Only the active adapter changes, which changes how the prompt is built and how the reply is decoded.

Which to use

Adapter	Best for
ChatAdapter	The default. Any model, any signature.
JSONAdapter	Rich output types (case classes, lists) on models with good JSON support.

For structured outputs, the output type is just a case class or enum that derives Schema, exactly as on the Signatures page. The adapter handles turning it into a prompt and back.

Next: Caching & usage.