OpenResponses

Mozaik's object model is designed to match OpenResponses, a shared way to describe LLM interactions across vendors: the same conceptual items for context and responses, even when wire formats differ.

Inputs and outputs

• Input (context) — Client-supplied items such as UserMessageItem, DeveloperMessageItem, SystemMessageItem, and FunctionCallOutputItem (tool results fed back to the model).
• Output (response) — Model-produced items such as ReasoningItem, FunctionCallItem, and ModelMessageItem.
• Streaming — Incremental provider output is exchanged as SemanticEvent<T> (type + data) on the environment before full items are assembled. See Streaming.

Mozaik turns that vocabulary into TypeScript classes you can construct, append to a ModelContext, serialize, and store.

Provider resolution

You select a model by its ModelName string when you call runInference. Mozaik resolves the name to a provider endpoint and model specification internally, maps the ModelContext to that provider's API, and returns typed ContextItems (and SemanticEvents when streaming is enabled).

Structured output uses the same ModelMessageItem shape — valid JSON in the text field, no new item type. See Structured output.

Why it matters

When your domain model follows a provider-agnostic spec:

• Swapping or combining providers becomes a smaller change — you change the ModelName, not your domain model.
• Persistence and replay stay stable — your database stores typed context items, not one-off JSON blobs per vendor.
• Tool use, reasoning traces, and assistant messages share one mental model across the whole environment.

Learn more

• OpenResponses — openresponses.org
• Mozaik repository