Vercel AI SDK + NocturnusAI

What changes in your API route

Call noct.processTurns() before streamText. Pass the briefingDelta as the system message. streamText receives only the new user message — never the full thread.

// app/api/chat/route.ts
import { streamText, type Message } from 'ai';
import { openai } from '@ai-sdk/openai';
import { NocturnusAIClient } from 'nocturnusai-sdk';

const noct = new NocturnusAIClient({
  baseUrl: process.env.NOCTURNUS_URL ?? 'http://localhost:9300',
  tenantId: 'default',
});

export async function POST(req: Request) {
  const { messages, sessionId } =
    (await req.json()) as { messages: Message[]; sessionId: string };
  const lastUserMessage = messages[messages.length - 1]?.content ?? '';

  // Compress prior context into a short briefing
  const ctx = await noct.processTurns({
    turns: [lastUserMessage],
    sessionId,
    scope: sessionId,
  });

  // streamText sees a short system prompt + the new user message only
  const result = streamText({
    model: openai('gpt-4o-mini'),
    system: ctx.briefingDelta ?? 'Start of conversation.',
    messages: [{ role: 'user', content: lastUserMessage }],
  });

  return result.toDataStreamResponse();
}

What changes in your page

Nothing. Standard useChat UI with a persistent sessionId body param.

// app/page.tsx
'use client';
import { useChat } from 'ai/react';
import { useMemo } from 'react';

export default function Chat() {
  const sessionId = useMemo(() => `session-${crypto.randomUUID()}`, []);
  const { messages, input, handleInputChange, handleSubmit } = useChat({
    api: '/api/chat',
    body: { sessionId },
  });
  // ...standard chat UI
}

Why it works

On a default Vercel AI SDK chat, streamText receives the full messages array every turn. By turn 10 the prompt is thousands of tokens of repeated history. With NocturnusAI, the prompt stays flat — the model sees only the briefingDelta for this turn.

What’s in the repo

• app/api/chat/route.ts — the compression handler
• app/page.tsx — standard useChat UI
• bench.mjs — measure your own tokens with node bench.mjs
• package.json — Next.js + ai + @ai-sdk/openai + nocturnusai-sdk