Use Attachments Skill

Process steps in order. Do not skip ahead.

Step 1 — Confirm the Model Supports Multimodal Input

Not all models accept attachments. Quick guide:

• OpenAI: GPT4o and later support images. Audio/PDF support varies by exact model
• Anthropic: Opus_4_*, Sonnet_4_* accept images and PDFs
• Google: Gemini_2.5_* accept images, video, audio
• Ollama / local models: depends on the specific model (e.g., LLaVA, Llama 3.2 Vision)

If the user's chosen model doesn't support the attachment type they want, redirect to a supporting model in the same provider before continuing.

Proceed immediately to Step 2.

Step 2 — Attach Content in the Prompt DSL

The 1.0 prompt DSL exposes an attachments block on user turns:

import ai.koog.prompt.dsl.prompt
import java.io.File

val visionPrompt = prompt("describe-screenshot") {
    user(
        text = "What's wrong with this UI?",
        attachments = listOf(
            Attachment.image(File("/path/to/screenshot.png")),
        ),
    )
}

Attachment factories cover the common cases:

• Attachment.image(file) / Attachment.image(url) / Attachment.image(bytes, mimeType)
• Attachment.file(file) — for PDF/document support on providers that accept them
• Attachment.audio(file) — for audio-capable models

Koog handles provider-specific encoding (base64 inlining vs URL reference vs uploaded-blob references) — you pass the file/bytes/URL, the executor adapts to the provider's wire shape.

Proceed immediately to Step 3.

Step 3 — Use Attachments Inside a Strategy

When attachments come from runtime input (uploads from a Ktor endpoint, file paths from CLI args), build them inside a node body and append to the prompt via llm.writeSession:

val strategy = strategy<File, String>("describe-image") {
    val describe by node<File, Message.User>("build-message") { imageFile ->
        Message.User(
            content = "Describe this image in detail.",
            attachments = listOf(Attachment.image(imageFile)),
        )
    }

    val ask by nodeLLMSendMessage()

    edge(nodeStart forwardTo describe)
    edge(describe forwardTo ask)
    edge(ask forwardTo nodeFinish onTextMessage { true })
}

For large attachments, prefer URL-based references over inline bytes — base64 inlining inflates request size and counts against token budgets (see add-token-budgeting).

Reference example: examples/simple-examples/.../attachments/ in the repo.

Finish here.