ai-generation-persistence（AI 一代的坚持）

AI Generation Persistence

AI generations are expensive, non-reproducible assets. Never discard them.

Every call to an LLM costs real money and produces unique output that cannot be exactly reproduced. Treat generations like database records — assign an ID, persist immediately, and make them retrievable.

Core Rules

1. Generate an ID before the LLM call — use nanoid() or createId() from @paralleldrive/cuid2
2. Persist every generation — text and metadata to database, images and files to Vercel Blob
3. Make every generation addressable — URL pattern: /chat/[id], /generate/[id], /image/[id]
4. Track metadata — model name, token usage, estimated cost, timestamp, user ID
5. Never stream without saving — if the user refreshes, the generation must survive

Generate-Then-Redirect Pattern

The standard UX flow for AI features: create the resource first, then redirect to its page.

// app/api/chat/route.ts
import { nanoid } from "nanoid";
import { db } from "@/lib/db";
import { redirect } from "next/navigation";

export async function POST(req: Request) {
  const { prompt, model } = await req.json();
  const id = nanoid();

  // Create the record BEFORE generation starts
  await db.insert(generations).values({
    id,
    prompt,
    model,
    status: "pending",
    createdAt: new Date(),
  });

  // Redirect to the generation page — it handles streaming
  redirect(`/chat/${id}`);
}

// app/chat/[id]/page.tsx
import { db } from "@/lib/db";
import { notFound } from "next/navigation";

export default async function ChatPage({ params }: { params: Promise<{ id: string }> }) {
  const { id } = await params;
  const generation = await db.query.generations.findFirst({
    where: eq(generations.id, id),
  });
  if (!generation) notFound();

  // Render with streaming if still pending, or show saved result
  return <ChatView generation={generation} />;
}

This gives you: shareable URLs, back-button support, multi-tab sessions, and generation history for free.

Persistence Schema

// lib/db/schema.ts
import { pgTable, text, integer, timestamp, jsonb } from "drizzle-orm/pg-core";

export const generations = pgTable("generations", {
  id: text("id").primaryKey(),            // nanoid
  userId: text("user_id"),                // auth user
  model: text("model").notNull(),         // "openai/gpt-5.4"
  prompt: text("prompt"),                 // input text
  result: text("result"),                 // generated output
  imageUrls: jsonb("image_urls"),         // Blob URLs for generated images
  tokenUsage: jsonb("token_usage"),       // { promptTokens, completionTokens }
  estimatedCostCents: integer("estimated_cost_cents"),
  status: text("status").default("pending"), // pending | streaming | complete | error
  createdAt: timestamp("created_at").defaultNow(),
});

Storage Strategy

Image Persistence

Never serve generated images as ephemeral base64 or temporary URLs. Save to Blob immediately:

import { put } from "@vercel/blob";
import { generateText } from "ai";

const result = await generateText({ model, prompt });

// Save every generated image to permanent storage
const imageUrls: string[] = [];
for (const file of result.files ?? []) {
  if (file.mediaType?.startsWith("image/")) {
    const ext = file.mediaType.split("/")[1] || "png";
    const blob = await put(`generations/${generationId}.${ext}`, file.uint8Array, {
      access: "public",
      contentType: file.mediaType,
    });
    imageUrls.push(blob.url);
  }
}

// Update the generation record with permanent URLs
await db.update(generations)
  .set({ imageUrls, status: "complete" })
  .where(eq(generations.id, generationId));

Cost Tracking

Extract usage from every generation and store it. This enables billing, budgeting, and abuse detection:

const result = await generateText({ model, prompt });

const usage = result.usage; // { promptTokens, completionTokens, totalTokens }
const estimatedCostCents = estimateCost(model, usage);

await db.update(generations).set({
  result: result.text,
  tokenUsage: usage,
  estimatedCostCents,
  status: "complete",
}).where(eq(generations.id, generationId));

Prompt Dedup / Caching

Avoid paying for identical generations. Cache by content hash:

import { Redis } from "@upstash/redis";
import { createHash } from "crypto";

const redis = Redis.fromEnv();

function hashPrompt(model: string, prompt: string): string {
  return createHash("sha256").update(`${model}:${prompt}`).digest("hex");
}

// Check cache before generating
const cacheKey = `gen:${hashPrompt(model, prompt)}`;
const cached = await redis.get<string>(cacheKey);
if (cached) return cached; // Return cached generation ID

// After generation, cache the result
await redis.set(cacheKey, generationId, { ex: 3600 }); // 1hr TTL

Anti-Patterns

• Streaming to client without saving — generation lost on page refresh. Always write to DB as tokens arrive or on completion.
• Routes without [id] segments — /api/chat with no ID means generations aren't addressable. Use /chat/[id].
• Re-generating identical prompts — check cache first. Same prompt + same model = same cost for no new value.
• Ephemeral base64 images — generated images served inline are lost when the component unmounts. Save to Vercel Blob.
• Missing metadata — always store model name, token counts, and timestamp. You need this for cost tracking and debugging.
• Client-only state — storing generations only in React state or localStorage. Use a database — generations must survive across devices and sessions.