Contents

Instructions

Details you provide that define the model’s intended behavior on prompts.

Declaration

struct Instructions

Mentioned in

Overview

Instructions are typically provided by you to define the role and behavior of the model. In the code below, the instructions specify that the model replies with topics rather than, for example, a recipe:

let instructions = """
    Suggest related topics. Keep them concise (three to seven words) and make sure they \
    build naturally from the person's topic.
    """

let session = LanguageModelSession(instructions: instructions)

let prompt = "Making homemade bread"
let response = try await session.respond(to: prompt)

Apple trains the model to obey instructions over any commands it receives in prompts, so don’t include untrusted content in instructions. For more on how instructions impact generation quality and safety, see Improving the safety of generative model output.

All input to the model contributes tokens to the context window of the LanguageModelSession — including the Instructions, Prompt, Tool, and Generable types, and the model’s responses. If your session exceeds the available context size, it throws LanguageModelSession.GenerationError.exceededContextWindowSize(_:).

Instructions can consume a lot of tokens that contribute to the context window size. To reduce your instruction size:

  • Write shorter instructions to save tokens.

  • Provide only the information necessary to perform the task.

  • Use concise and imperative language instead of indirect or jargon that the model might misinterpret.

  • Aim for one to three paragraphs instead of including a significant amount of background information, policy, or extra content.

For more information on managing the context window size, see TN3193: Managing the on-device foundation model’s context window.

Topics

Creating instructions

See Also

Prompting