feat(gemini): Skiva 1 – Gemini 2.5 Flash som lärar-tier för kylskåpsskanning
- Gemini-adapter bakom AamosClient-interface (AAMOS_MODE=gemini) - Serversida/worker: hämtar bild, anropar Gemini, mappar mot canonical_ingredients - Kostnad/tokens bokförs i ai_usage_counters; global dagsbudget via BudgetStore - Redis-backed budget i worker, in-memory i tester - Migration 0019: ai_cost_usd_microcents - Hermetiska tester med inspelad fixture; separat pnpm eval:scan - docs/09 uppdaterad ärligt: AAMOS-status, Gemini-flöde, säkerhet/kostnad - REQUIRE_REAL=1 stödjer AAMOS_MODE=gemini; deploy-grind uppdaterad
This commit is contained in:
@@ -0,0 +1,162 @@
|
||||
import { config as loadDotenv } from "dotenv";
|
||||
import { existsSync } from "node:fs";
|
||||
import path from "node:path";
|
||||
for (const candidate of [".env", "../.env", "../../.env"]) {
|
||||
const p = path.resolve(process.cwd(), candidate);
|
||||
if (existsSync(p)) {
|
||||
loadDotenv({ path: p });
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
import { GeminiAamosClient, type BudgetStore } from "@app/ai-contracts";
|
||||
|
||||
/**
|
||||
* Live-evaluering av Gemini-lärar-tier för kylskåps-/skafferiskanning.
|
||||
*
|
||||
* pnpm --filter @app/worker eval:scan
|
||||
*
|
||||
* Kräver AAMOS_MODE=gemini + GEMINI_API_KEY. Använder ett litet guldset av
|
||||
* publikt tillgängliga bild-URL:er (eller lokala filer). Rapporterar
|
||||
* träffbild, kvantitet, enhet, kostnad och latens. Avslutar med exitkod 1
|
||||
* om någon obligatorisk kontroll fallerar.
|
||||
*/
|
||||
|
||||
interface ScanEvalCase {
|
||||
id: string;
|
||||
imageUrls: string[];
|
||||
locationType: "fridge" | "pantry";
|
||||
marketLocale: string;
|
||||
checks: (items: Array<{
|
||||
detectedName: string;
|
||||
brand: string | null;
|
||||
estimatedQuantity: number | null;
|
||||
unit: string | null;
|
||||
confidence: number;
|
||||
requiresConfirmation: boolean;
|
||||
}>) => { name: string; passed: boolean }[];
|
||||
}
|
||||
|
||||
const CASES: ScanEvalCase[] = [
|
||||
{
|
||||
id: "arla-milk-fridge",
|
||||
imageUrls: ["https://upload.wikimedia.org/wikipedia/commons/6/6c/Arla_Ko_Mellanmj%C3%B6lk_1L.jpg"],
|
||||
locationType: "fridge",
|
||||
marketLocale: "sv-SE",
|
||||
checks: (items) => {
|
||||
const milk = items.find((i) =>
|
||||
/mjölk|milk/i.test(i.detectedName) || /arla/i.test(i.brand ?? ""),
|
||||
);
|
||||
return [
|
||||
{ name: "hittade mjölkprodukt", passed: !!milk },
|
||||
{ name: "konfidens > 0.7", passed: !!milk && milk.confidence > 0.7 },
|
||||
{ name: "kräver bekräftelse om låg konfidens", passed: !!milk && (milk.confidence >= 0.92 || milk.requiresConfirmation) },
|
||||
];
|
||||
},
|
||||
},
|
||||
{
|
||||
id: "swedish-butter-fridge",
|
||||
imageUrls: ["https://upload.wikimedia.org/wikipedia/commons/3/3e/Svenskt_Sm%C3%B6r_Normalsaltat_80-25_500g.jpg"],
|
||||
locationType: "fridge",
|
||||
marketLocale: "sv-SE",
|
||||
checks: (items) => {
|
||||
const butter = items.find((i) => /smör|butter/i.test(i.detectedName));
|
||||
return [
|
||||
{ name: "hittade smör", passed: !!butter },
|
||||
{ name: "konfidens > 0.7", passed: !!butter && butter.confidence > 0.7 },
|
||||
];
|
||||
},
|
||||
},
|
||||
];
|
||||
|
||||
class SilentBudgetStore implements BudgetStore {
|
||||
private spend = 0;
|
||||
async getDailySpendUsd(): Promise<number> {
|
||||
return this.spend;
|
||||
}
|
||||
async incrementDailySpendUsd(amountUsd: number): Promise<number> {
|
||||
this.spend += amountUsd;
|
||||
return this.spend;
|
||||
}
|
||||
}
|
||||
|
||||
async function main() {
|
||||
const mode = process.env.AAMOS_MODE ?? "mock";
|
||||
const apiKey = process.env.GEMINI_API_KEY;
|
||||
|
||||
if (mode !== "gemini" || !apiKey) {
|
||||
console.error("[eval:scan] KRÄVER AAMOS_MODE=gemini + GEMINI_API_KEY.");
|
||||
process.exit(1);
|
||||
}
|
||||
|
||||
const client = new GeminiAamosClient({
|
||||
apiKey,
|
||||
model: process.env.GEMINI_MODEL,
|
||||
timeoutMs: process.env.GEMINI_TIMEOUT_MS ? Number(process.env.GEMINI_TIMEOUT_MS) : 60_000,
|
||||
dailyBudgetUsd: process.env.GEMINI_DAILY_BUDGET_USD ? Number(process.env.GEMINI_DAILY_BUDGET_USD) : 10,
|
||||
budgetStore: new SilentBudgetStore(),
|
||||
});
|
||||
|
||||
console.log(`[eval:scan] Gemini-lärar-tier live-evaluering – ${CASES.length} fall\n`);
|
||||
|
||||
let totalChecks = 0;
|
||||
let failedChecks = 0;
|
||||
let totalCostUsd = 0;
|
||||
let totalLatencyMs = 0;
|
||||
|
||||
for (const evalCase of CASES) {
|
||||
const started = Date.now();
|
||||
let checks: { name: string; passed: boolean }[] = [];
|
||||
let error: string | null = null;
|
||||
|
||||
try {
|
||||
const result = await client.runTask(
|
||||
evalCase.locationType === "fridge" ? "ANALYZE_FRIDGE_IMAGE" : "ANALYZE_PANTRY_IMAGE",
|
||||
{
|
||||
imageUrls: evalCase.imageUrls,
|
||||
locationType: evalCase.locationType,
|
||||
marketLocale: evalCase.marketLocale,
|
||||
knownItems: [],
|
||||
},
|
||||
);
|
||||
|
||||
if (result.status !== "ok" || !result.output) {
|
||||
error = `status=${result.status}: ${result.error ?? "okänt fel"}`;
|
||||
checks = [{ name: "Gemini svarade ok", passed: false }];
|
||||
} else {
|
||||
const items = result.output.items;
|
||||
checks = [{ name: "Gemini svarade ok", passed: true }, ...evalCase.checks(items)];
|
||||
}
|
||||
|
||||
if (result.costUsd) totalCostUsd += result.costUsd;
|
||||
if (result.latencyMs) totalLatencyMs += result.latencyMs;
|
||||
} catch (err) {
|
||||
error = (err as Error).message;
|
||||
checks = [{ name: "Gemini svarade ok", passed: false }];
|
||||
}
|
||||
|
||||
const ms = Date.now() - started;
|
||||
const passed = checks.every((c) => c.passed);
|
||||
totalChecks += checks.length;
|
||||
failedChecks += checks.filter((c) => !c.passed).length;
|
||||
|
||||
console.log(`${passed ? "✓" : "✗"} ${evalCase.id} (${ms} ms)`);
|
||||
for (const c of checks) if (!c.passed) console.log(` ✗ ${c.name}`);
|
||||
if (error) console.log(` fel: ${error}`);
|
||||
}
|
||||
|
||||
console.log(`\n[eval:scan] ${CASES.length} fall, ${totalChecks} kontroller, ${failedChecks} fallerade.`);
|
||||
console.log(`[eval:scan] Total latens: ${totalLatencyMs} ms, total kostnad: ~$${totalCostUsd.toFixed(6)}`);
|
||||
|
||||
if (failedChecks > 0) {
|
||||
console.log("[eval:scan] UNDERKÄND – åtgärda innan Skiva 1 går till fälttest.");
|
||||
process.exit(1);
|
||||
}
|
||||
console.log("[eval:scan] GODKÄND – fälttest kan börja.");
|
||||
process.exit(0);
|
||||
}
|
||||
|
||||
main().catch((err) => {
|
||||
console.error("[eval:scan] KRASCH:", err);
|
||||
process.exit(1);
|
||||
});
|
||||
Reference in New Issue
Block a user