Data types

Pydantic models used by the SDK – message hierarchy, token usage, tool calls and more.

All public result objects & helper parameters rely on a small set of immutable Pydantic models defined in ai_sdk.types.

---

Messages

CoreMessage subclasses

Class	Role	Typical usage
CoreSystemMessage	system	Set behaviour & identity of the assistant.
CoreUserMessage	user	Explicit user instructions (text / images / files).
CoreAssistantMessage	assistant	Model responses (text, reasoning, tool calls).
CoreToolMessage	tool	Container for tool results fed back to the model.

Each subclass exposes a .to_dict() helper that returns the OpenAI-compatible representation, so you can seamlessly drop them into the low-level LanguageModel interface if needed.

Content parts

parts_breakdown.py

from ai_sdk.types import TextPart, ImagePart, ToolCallPart, ToolResultPart

Parts make multimodal messages explicit. For example a user message containing both text and an image:

CoreUserMessage(
content=[
  TextPart(text="Describe this painting"),
  ImagePart(image="https://example.com/monet.jpg", mime_type="image/jpeg"),
]
)

---

TokenUsage

Simple container holding prompt_tokens, completion_tokens, total_tokens – populated on all helpers when the provider returns usage data.

ToolCall / ToolResult

Used internally for the tool-calling loop but exposed on GenerateTextResult so you can inspect or persist the chain-of-thought.

OnStepFinishResult

Object passed to the optional on_step callback after every model ↔ tool interaction. Fields include:

• step_type – one of initial, continue, tool-result.
• finish_reason, usage, text, tool_calls, tool_results, …

---

All types are frozen (pydantic.ConfigDict(frozen=True)) meaning you can safely log & share them without accidental mutation.