@langinsight/ai-sdk

LangInsight callback handler for Vercel AI SDK. Sends input/output traces to LangInsight in parallel and supports feedback (thumbs up/down) via score().

Usage

generateText

import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
import { LangInsight } from "@langinsight/ai-sdk";

const { LANGINSIGHT_ENDPOINT, LANGINSIGHT_API_KEY } = process.env;

const handler = new LangInsight.CallbackHandler({
  apiKey: LANGINSIGHT_API_KEY,
  endpoint: LANGINSIGHT_ENDPOINT,
  metadata: { userId: "admin", sessionId: "admin" },
  onSuccess: (input, output) => {
    // traceId の取得・DB永続化はここで
    console.log("Trace sent:", output.id);
  },
});

const prompt = "Hello!";

const result = await generateText({
  model: openai("gpt-4o"),
  prompt,
});

await handler.report(result, { input: prompt });

streamText

import { streamText } from "ai";
import { openai } from "@ai-sdk/openai";
import { LangInsight } from "@langinsight/ai-sdk";

const { LANGINSIGHT_ENDPOINT, LANGINSIGHT_API_KEY } = process.env;

const handler = new LangInsight.CallbackHandler({
  apiKey: LANGINSIGHT_API_KEY,
  endpoint: LANGINSIGHT_ENDPOINT,
  metadata: { userId: "admin", sessionId: "admin" },
  onSuccess: (input, output) => console.log("Trace sent:", output.id),
  onFailure: (err) => console.error("Trace failed:", err),
});

const prompt = "Hello!";

const result = streamText({
  model: openai("gpt-4o"),
  prompt,
  onFinish: handler.onFinish({ input: prompt }),
});

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

会話ループ（Ollama でダミー trace を生成）

Ollama 同士で会話を繰り返し、各ターンの input/output を plugin 経由で Core API (trace) に送る example。

cd plugins/typescript/ai-sdk
LANGINSIGHT_ENDPOINT=http://localhost:3000 LANGINSIGHT_API_KEY=xxx bun run examples/conversation-loop.ts
# オプション: TURNS=5（デフォルト5）, SEED_PROMPT=最初のユーザー発話

各ターンで handler.report(result) が呼ばれ、trace が Core API に送信される。

Feedback (score)

Trace ID は handler.result で取得し、その output.id を score() に渡してフィードバックを送信する。DB 永続化などは onSuccess コールバックでも行える。

const handler = new LangInsight.CallbackHandler({
  ...options,
});

// report() または onFinish の後に
const { input, output } = await handler.result;
await handler.score({ traceId: output.id, value: 1 });  // 👍
await handler.score({ traceId: output.id, value: -1 }); // 👎

Options

| Option | Type | Required | Description | | ----------- | --------------------------------- | -------- | --------------------------------------------------------------------------- | | apiKey | string | Yes | LangInsight API key | | endpoint | string | Yes | LangInsight API endpoint (e.g. https://api.langinsight.example.com) | | metadata | object | Yes | Metadata attached to traces | | metadata.userId | string | Yes | User ID | | metadata.sessionId| string | Yes | Session ID (e.g. conversation id) | | metadata.modelName | string | No | Model name (defaults to result response when not set) | | metadata.* | any | No | Additional key-value pairs | | onSuccess | (input: Trace, output: Trace) => void | No | Callback invoked with both input and output traces when trace submission succeeds. Use output.id to get traceId and persist to DB. | | onFailure | (error: Error) => void | No | Callback invoked with the error when trace submission fails |

API

• report(result, { input }) — Sends traces for a generateText result. Both input and output traces are passed to onSuccess on success.
• onFinish({ input }) — Returns a function suitable for streamText's onFinish. Sends traces when the stream finishes; both traces are passed to onSuccess.
• score({ traceId, value }) — Submits feedback (1 = thumbs up, -1 = thumbs down) for the given trace. Use (await handler.result).output.id as traceId.
• result — A promise that resolves with { input: Trace, output: Trace } when traces are sent successfully.