genkit-aisdk-converter

Convert between Genkit generateStream output and the AI SDK UI message stream protocol. Use Genkit on the server for generation while staying compatible with AI SDK clients (useChat, AIChatAgent, etc.).

Overview

• Incoming (client → server): AI SDK UI messages → Genkit MessageData[] so your endpoint can pass them into Genkit.
• Outgoing (server → client): Genkit GenerateResponseChunk stream → AI SDK UI message stream so the same endpoint can return a Response that useChat() and other AI SDK clients consume.

Installation

npm install genkit-aisdk-converter

Peer dependencies: ai (≥6.0.0) and @genkit-ai/ai (≥1.0.0). Install them if not already present:

npm install ai @genkit-ai/ai

Usage

Outgoing: Genkit stream → AI SDK response

Turn a Genkit generateStream() result into a streaming Response for useChat():

import { ai } from "@genkit-ai/ai";
import {
  genkitStreamToUIMessageStreamResponse,
  uiMessagesToGenkitMessages,
} from "genkit-aisdk-converter";

// In your route/handler:
const genkitStream = await ai.generateStream({
  model: "your-model",
  messages: uiMessagesToGenkitMessages(messages), // convert client messages
});

const response = genkitStreamToUIMessageStreamResponse(
  genkitStream
);

return response;

To get a ReadableStream<UIMessageChunk> instead of a Response (e.g. to wrap with createUIMessageStreamResponse() yourself):

import { genkitChunksToUIMessageStream } from "genkit-aisdk-converter";
import { createUIMessageStreamResponse } from "ai";

const stream = genkitChunksToUIMessageStream(genkitStream, options);
return createUIMessageStreamResponse(stream);

Incoming: AI SDK messages → Genkit messages

Convert messages from useChat (or AIChatAgent) to Genkit MessageData[] before calling ai.generateStream():

import { uiMessagesToGenkitMessages } from "genkit-aisdk-converter";

const genkitMessages = uiMessagesToGenkitMessages(messages, {
  ignoreIncompleteToolCalls: true, // omit tool parts still streaming input
  convertDataPart: (part) => {
    // Optional: AI SDK "data" parts (e.g. type "data-custom") are arbitrary JSON.
    // Genkit only accepts text/media, so by default they are skipped. Return
    // { text } or { media } to include them in the Genkit message.
    if (part.data?.type === "caption") return { text: part.data.text };
    return undefined;
  },
});

Handled UI part types: text, file, reasoning, tool (with ref/name/input/output and approval/denied), data (with optional convertDataPart), step-start, source-url, source-document.

API

| Export | Description | |--------|-------------| | uiMessagesToGenkitMessages(messages, options?) | AI SDK UIMessage[] → Genkit MessageData[]. | | genkitChunksToUIMessageStream(genkitStream, options?) | AsyncIterable<GenerateResponseChunk> → ReadableStream<UIMessageChunk>. | | genkitStreamToUIMessageStreamResponse(genkitStream, init?, options?) | Same stream converted into a streaming Response for AI SDK clients. |

Options

UiMessagesToGenkitOptions (incoming)

• convertDataPart?(part) – AI SDK "data" parts (e.g. data-custom) are arbitrary JSON and are skipped by default. Return { text } or { media } to include them in Genkit messages; return undefined to keep skipping.
• ignoreIncompleteToolCalls? – When true, omit tool parts in input-streaming / input-available (aligns with convertToModelMessages).

GenkitStreamToUIMessageOptions (outgoing)

• responsePromise? – e.g. genkitStream.response; when provided, the real Genkit finish reason is used on the final finish chunk (otherwise it defaults to "stop").
• messageMetadata? – Attached to start and finish chunks (e.g. for tracing).
• dataPartTypeName? – Name for custom/data chunks (default 'custom'); alphanumeric and hyphen only; produces data-<name>.
• toolResponseToChunk? – Optional mapper to emit tool-output-error or tool-output-denied instead of tool-output-available.

Protocol mapping

• Genkit chunks: { index, role, content: Part[] } with Part = { text } | { toolRequest } | { reasoning } | { media } | custom, etc.
• AI SDK UI stream: start → text-start / text-delta / text-end, reasoning-start / reasoning-delta / reasoning-end, tool-input-start / tool-input-available / tool-output-available, finish-step, finish → [DONE].

License

MIT · Invertase