​

Guide · OpenAI SDK

​

1. Stack

• Runtime: Next.js App Router (Route Handlers) or Edge Functions.
• SDK: openai official client + ai (Vercel AI SDK) for streaming.
• Language: TypeScript with strict types.

​

2. Boilerplate

import OpenAI from "openai";
import { StreamingTextResponse, experimental_StreamData } from "ai";

const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

export async function POST(req: Request) {
  const { prompt } = await req.json();
  const response = await openai.responses.create({
    model: "gpt-4o-mini",
    input: prompt,
    temperature: 0.2,
  });
  return new Response(response.output[0].content[0].text);
}

​

3. Patterns

• Function calling: define tool schema with Zod; handle response.output[0].content accordingly.
• Vision: send base64 images or remote URLs for document QA.
• Batching: use the Batch API for large eval jobs; monitor status webhooks.

​

4. Cost Controls

• Use gpt-4o-mini for drafts, gpt-4o for final outputs.
• Log tokens per request; feed into cost observability dashboard.
• Cache responses (KV store) for deterministic prompts (e.g., onboarding instructions).

​

5. Safety

• Apply moderation endpoint before returning user-facing output.
• Combine with Guardrails (structured outputs + policy checks).

​

6. Testing

• Build fixtures using the Responses API seed parameter for repeatability.
• Run contract tests in CI hitting OpenAI mock server (e.g., msw).