Adapters: chat markers or JSON

The adapter decides the prompt format and how replies are parsed. ChatAdapter uses field markers; JSONAdapter asks for a JSON object.

Example

inputs = {"ticket": "My parcel never arrived"}
for adapter in [dspy.ChatAdapter(), dspy.JSONAdapter()]:
    messages = adapter.format(Triage, demos=[], inputs=inputs)
    print(f"--- {type(adapter).__name__}")
    print(messages[-1]["content"])

The signature and inputs are unchanged; only the instructions for the answer differ. JSONAdapter asks for a JSON object with the output fields as keys, and it asks providers that support it to enforce JSON output.

Example

print(dspy.JSONAdapter().parse(Triage, '{"category": "shipping"}'))

When a reply cannot be parsed

Example

with dspy.context(adapter=dspy.JSONAdapter()):
    dspy.Predict(Triage)(ticket="My parcel never arrived")

dspy.context(adapter=...) switches the adapter like it switched the model. The stand-in only writes markers, so JSONAdapter cannot parse its reply, and the error includes the reply it got. A real model follows whichever format the prompt asks for.

The opposite direction is handled for you: when ChatAdapter cannot parse a reply, it retries the request once with JSONAdapter before raising. Models that ignore the markers often manage JSON.

Adapter	Use it when
`ChatAdapter`	The default. Works with any chat model.
`JSONAdapter`	The provider supports structured output, or answers have nested types.
`XMLAdapter`	The model is trained to read and write XML tags.
`TwoStepAdapter`	A reasoning model writes freely and a second, cheaper model extracts the fields.

Try it yourself

Format Triage with dspy.XMLAdapter() and print the user message.
Parse '{"category": "refunds"}' with JSONAdapter.
Configure adapter=dspy.JSONAdapter() globally and run the mock from lesson 2.

PreviousYour own modules: steps in plain Python

This is what real progress feels like.