feat(worker/ai): Gemini-lärar-tier för scan/detect/allergen + shadow-capture

- Utöka GeminiAamosClient med ANALYZE_MEAL_IMAGE, READ_NUTRITION_LABEL,
  READ_EXPIRY_DATE och READ_RECEIPT (vision-prompts mot kontraktsscheman).
- Lägg till apps/worker/src/lib/shadow-capture.ts: S3/local fallback,
  imageTraining-gating, ingen PII, aldrig faila användarvägen.
- Integrera capture i scan-processorn efter lyckad runTask med consentFlags.
- Lägg till @aws-sdk/client-s3 i worker samt eval:capture-skript.
- Uppdatera eval:scan till Wikimedia Special:FilePath + redirect-following.
- Exkludera .terraform i brand-guard för att undvika false positives.
- Uppdatera gemini-test för DEDUPLICATE_INVENTORY som unsupported.
This commit is contained in:
Sven (AAMOS AI)
2026-08-12 02:51:05 +07:00
parent 515dd12cc7
commit 9a9f8a6a3c
9 changed files with 615 additions and 20 deletions
+341 -7
View File
@@ -1,5 +1,5 @@
/**
* Gemini-lärar-tier för AAMOS-task-typer (SKIVA 1: kylskåpsskanning).
* Gemini-lärar-tier för AAMOS-task-typer (SKIVA 1).
*
* - Implementerar samma AamosClient-interface som HttpAamosClient/MockAamosClient.
* - App + Food API ser ingen skillnad; workern byter bara adapter.
@@ -7,7 +7,6 @@
* - Kostnad/tokens rapporteras tillbaka så workern kan bokföra i ai_usage_counters.
* - Global dagsbudget kan sättas via GEMINI_DAILY_BUDGET_USD.
*/
import { randomUUID } from "node:crypto";
import { z } from "zod";
import {
TASK_CONTRACTS,
@@ -17,14 +16,22 @@ import {
detectedItemSchema,
type DetectedItem,
generateRecipeCandidatesOutput,
analyzeMealImageOutput,
readNutritionLabelOutput,
readExpiryDateOutput,
readReceiptOutput,
} from "./tasks.js";
import type { AamosCallOptions, AamosClient, AamosResult } from "./client.js";
import type { LocaleContext } from "@app/shared-types";
const GEMINI_API_BASE = "https://generativelanguage.googleapis.com/v1beta";
/** Custom fetch using Node https module — works around IPv6 hangs on WSL. */
/** Custom fetch using Node https module — works around IPv6 hangs on WSL. Follows redirects. */
function customNodeFetch(input: string | URL | Request, init?: RequestInit): Promise<Response> {
return requestOnce(input, init, 5);
}
function requestOnce(input: string | URL | Request, init: RequestInit | undefined, redirectsLeft: number): Promise<Response> {
return new Promise((resolve, reject) => {
const url = typeof input === "string" ? input : input instanceof URL ? input.toString() : input.url;
const u = new URL(url);
@@ -45,6 +52,9 @@ function customNodeFetch(input: string | URL | Request, init?: RequestInit): Pro
if (postData && !headers["content-length"]) {
headers["content-length"] = String(Buffer.byteLength(postData));
}
if (!headers["user-agent"]) {
headers["user-agent"] = "Mozilla/5.0 (compatible; Cibello/1.0)";
}
const req = httpMod.request(
{
@@ -56,13 +66,23 @@ function customNodeFetch(input: string | URL | Request, init?: RequestInit): Pro
timeout: 120_000,
},
(res) => {
const status = res.statusCode ?? 200;
const location = res.headers.location;
if (status >= 300 && status < 400 && location && redirectsLeft > 0) {
const nextUrl = new URL(location, url).toString();
requestOnce(nextUrl, { ...init, method: "GET", body: undefined }, redirectsLeft - 1)
.then(resolve)
.catch(reject);
return;
}
let body = "";
res.setEncoding("utf8");
res.on("data", (chunk) => { body += chunk; });
res.on("end", () => {
resolve(
new Response(body, {
status: res.statusCode ?? 200,
status,
statusText: res.statusMessage ?? "OK",
headers: new Headers(Object.entries(res.headers).map(([k, v]) => [k, String(v)])),
}),
@@ -81,6 +101,9 @@ function customNodeFetch(input: string | URL | Request, init?: RequestInit): Pro
/** Estimated max cost per image call in USD (pessimistic). */
const COST_ESTIMATE_PER_IMAGE_USD = 0.0015;
/** Estimated max cost per simple OCR image call in USD. */
const COST_ESTIMATE_PER_OCR_IMAGE_USD = 0.001;
export interface BudgetStore {
/** Return current daily spend in USD. */
getDailySpendUsd(): Promise<number>;
@@ -222,7 +245,7 @@ export class GeminiAamosClient implements AamosClient {
async healthCheck(): Promise<{ ok: boolean; detail?: string }> {
try {
const url = `${GEMINI_API_BASE}/models?key=${this.cfg.apiKey}`;
const url = `${GEMINI_API_BASE}/models?key=${encodeURIComponent(this.cfg.apiKey)}`;
const res = await this.cfg.fetchImpl(url, { signal: AbortSignal.timeout(5_000) });
return { ok: res.ok, detail: `status ${res.status}` };
} catch (err) {
@@ -251,6 +274,14 @@ export class GeminiAamosClient implements AamosClient {
return this.analyzeStorageImage(taskType, parsedInput.data as TaskInput<"ANALYZE_FRIDGE_IMAGE">, options) as Promise<AamosResult<T>>;
case "GENERATE_RECIPE_CANDIDATES":
return this.generateRecipeCandidates(parsedInput.data as TaskInput<"GENERATE_RECIPE_CANDIDATES">, options) as Promise<AamosResult<T>>;
case "ANALYZE_MEAL_IMAGE":
return this.analyzeMealImage(parsedInput.data as TaskInput<"ANALYZE_MEAL_IMAGE">, options) as Promise<AamosResult<T>>;
case "READ_NUTRITION_LABEL":
return this.readNutritionLabel(parsedInput.data as TaskInput<"READ_NUTRITION_LABEL">, options) as Promise<AamosResult<T>>;
case "READ_EXPIRY_DATE":
return this.readExpiryDate(parsedInput.data as TaskInput<"READ_EXPIRY_DATE">, options) as Promise<AamosResult<T>>;
case "READ_RECEIPT":
return this.readReceipt(parsedInput.data as TaskInput<"READ_RECEIPT">, options) as Promise<AamosResult<T>>;
default:
return {
status: "failed",
@@ -295,7 +326,7 @@ export class GeminiAamosClient implements AamosClient {
},
};
const url = `${GEMINI_API_BASE}/models/${this.cfg.model}:generateContent?key=${this.cfg.apiKey}`;
const url = `${GEMINI_API_BASE}/models/${this.cfg.model}:generateContent?key=${encodeURIComponent(this.cfg.apiKey)}`;
const res = await this.cfg.fetchImpl(url, {
method: "POST",
headers: { "content-type": "application/json" },
@@ -350,6 +381,186 @@ export class GeminiAamosClient implements AamosClient {
};
}
private async runVisionTask<T extends AamosTaskType>(
taskType: T,
imageUrls: string[],
prompt: string,
outputSchema: z.ZodSchema<unknown>,
mapOutput: (parsed: unknown) => TaskOutput<T>,
options: AamosCallOptions,
promptVersion: string,
costEstimatePerImage: number,
): Promise<AamosResult<T>> {
const locale = options.localeContext ?? this.defaultLocale();
const started = Date.now();
const imageParts = await this.fetchImageParts(imageUrls, this.cfg.timeoutMs);
const estimatedCostUsd = imageParts.length * costEstimatePerImage;
if (await this.isOverBudget(estimatedCostUsd)) {
return {
status: "failed",
output: null,
error: "Global Gemini-dagsbudget är förbrukad.",
} as AamosResult<T>;
}
const payload = {
contents: [{ parts: [{ text: prompt }, ...imageParts] }],
generationConfig: {
responseMimeType: "application/json",
temperature: 0.2,
},
};
const url = `${GEMINI_API_BASE}/models/${this.cfg.model}:generateContent?key=${encodeURIComponent(this.cfg.apiKey)}`;
const res = await this.cfg.fetchImpl(url, {
method: "POST",
headers: { "content-type": "application/json" },
body: JSON.stringify(payload),
signal: AbortSignal.timeout(this.cfg.timeoutMs),
});
if (!res.ok) {
const body = await res.text();
return {
status: "failed",
output: null,
error: `Gemini ${res.status}: ${body.slice(0, 500)}`,
} as AamosResult<T>;
}
const geminiBody = (await res.json()) as {
candidates?: Array<{ content?: { parts?: Array<{ text?: string }> } }>;
usageMetadata?: { promptTokenCount?: number; candidatesTokenCount?: number };
};
const text = geminiBody.candidates?.[0]?.content?.parts?.[0]?.text ?? "";
let parsed: unknown;
try {
const json = JSON.parse(text);
const safe = outputSchema.safeParse(json);
if (!safe.success) {
return {
status: "failed",
output: null,
error: `Gemini-svar matchar inte schema: ${safe.error.message}`,
} as AamosResult<T>;
}
parsed = safe.data;
} catch {
return {
status: "failed",
output: null,
error: "Gemini-svar var inte giltig JSON.",
} as AamosResult<T>;
}
const inputTokens = geminiBody.usageMetadata?.promptTokenCount ?? 0;
const outputTokens = geminiBody.usageMetadata?.candidatesTokenCount ?? 0;
const costUsd = this.estimateCostUsd(inputTokens, outputTokens);
await this.recordSpend(costUsd);
const latencyMs = Date.now() - started;
const output = mapOutput(parsed) as TaskOutput<T>;
return {
status: "ok",
output,
modelVersion: this.cfg.model,
promptVersion,
latencyMs,
costUsd,
inputTokens,
outputTokens,
};
}
private async analyzeMealImage(
input: TaskInput<"ANALYZE_MEAL_IMAGE">,
options: AamosCallOptions,
): Promise<AamosResult<"ANALYZE_MEAL_IMAGE">> {
const locale = options.localeContext ?? this.defaultLocale();
const prompt = this.buildMealPrompt(locale, input.recipeContext);
return this.runVisionTask(
"ANALYZE_MEAL_IMAGE",
input.imageUrls,
prompt,
analyzeMealImageOutput,
(parsed) => parsed as TaskOutput<"ANALYZE_MEAL_IMAGE">,
options,
"gemini-meal-v1",
COST_ESTIMATE_PER_IMAGE_USD,
);
}
private async readNutritionLabel(
input: TaskInput<"READ_NUTRITION_LABEL">,
options: AamosCallOptions,
): Promise<AamosResult<"READ_NUTRITION_LABEL">> {
const locale = options.localeContext ?? this.defaultLocale();
const prompt = this.buildNutritionPrompt(locale);
return this.runVisionTask(
"READ_NUTRITION_LABEL",
input.imageUrls,
prompt,
readNutritionLabelOutput,
(parsed) => parsed as TaskOutput<"READ_NUTRITION_LABEL">,
options,
"gemini-nutrition-v1",
COST_ESTIMATE_PER_OCR_IMAGE_USD,
);
}
private async readExpiryDate(
input: TaskInput<"READ_EXPIRY_DATE">,
options: AamosCallOptions,
): Promise<AamosResult<"READ_EXPIRY_DATE">> {
const locale = options.localeContext ?? this.defaultLocale();
const prompt = this.buildExpiryPrompt(locale);
return this.runVisionTask(
"READ_EXPIRY_DATE",
input.imageUrls,
prompt,
readExpiryDateOutput,
(parsed) => parsed as TaskOutput<"READ_EXPIRY_DATE">,
options,
"gemini-expiry-v1",
COST_ESTIMATE_PER_OCR_IMAGE_USD,
);
}
private async readReceipt(
input: TaskInput<"READ_RECEIPT">,
options: AamosCallOptions,
): Promise<AamosResult<"READ_RECEIPT">> {
const locale = options.localeContext ?? this.defaultLocale();
const prompt = this.buildReceiptPrompt(locale);
return this.runVisionTask(
"READ_RECEIPT",
input.imageUrls,
prompt,
readReceiptOutput,
(parsed) => parsed as TaskOutput<"READ_RECEIPT">,
options,
"gemini-receipt-v1",
COST_ESTIMATE_PER_IMAGE_USD,
);
}
private async fetchImageParts(
imageUrls: string[],
timeoutMs: number,
): Promise<Array<{ inline_data: { mime_type: string; data: string } }>> {
return Promise.all(
imageUrls.slice(0, 6).map(async (url) => {
const res = await this.cfg.fetchImpl(url, { signal: AbortSignal.timeout(timeoutMs) });
if (!res.ok) throw new Error(`Kunde inte hämta bild: ${res.status} ${url}`);
const buf = Buffer.from(await res.arrayBuffer());
return { inline_data: { mime_type: this.mimeType(buf), data: buf.toString("base64") } };
}),
);
}
private async generateRecipeCandidates(
input: TaskInput<"GENERATE_RECIPE_CANDIDATES">,
options: AamosCallOptions,
@@ -377,7 +588,7 @@ export class GeminiAamosClient implements AamosClient {
},
};
const url = `${GEMINI_API_BASE}/models/${this.cfg.model}:generateContent?key=${this.cfg.apiKey}`;
const url = `${GEMINI_API_BASE}/models/${this.cfg.model}:generateContent?key=${encodeURIComponent(this.cfg.apiKey)}`;
const res = await this.cfg.fetchImpl(url, {
method: "POST",
headers: { "content-type": "application/json" },
@@ -553,6 +764,129 @@ Svara ENDAST med giltig JSON i exakt detta format:
}
imageQualityIssues kan innehålla någon av: "dark", "blurry", "occlusion", "too_far".
Språk: ${lang}.`;
}
private buildMealPrompt(locale: LocaleContext, recipeContext: unknown): string {
const lang = locale.languageTag.startsWith("en") ? "English" : "Swedish";
const ctx = recipeContext as {
recipeId?: string;
titleSv?: string;
nutritionPerPortion?: Record<string, number>;
portions?: number;
} | null;
let contextLine = "";
if (ctx?.titleSv) {
contextLine = `Användaren säger att måltiden ska matcha receptet "${ctx.titleSv}" (${ctx.portions ?? "?"} portioner). `;
}
return `Du är en svensk näringsassistent. ${contextLine}Bilden visar en serverad måltid.
Uppskatta:
- matchesRecipeContext: true/false/null matchar bilden receptet användaren nämnde?
- portionFractionEstimate: hur stor del av en normalportion ser det ut att vara? 0.05.0, eller null.
- kcalRange: { min, max, mostLikely } för hela måltiden, eller null.
- components: lista med uppskattade komponenter (namn, ev. canonicalIngredientId som null, estimatedGrams, confidence).
- confidence: 0.01.0 för hela tolkningen.
Svara ENDAST med giltig JSON som matchar detta schema:
{
"matchesRecipeContext": boolean | null,
"portionFractionEstimate": number | null,
"kcalRange": { "min": number, "max": number, "mostLikely": number } | null,
"components": [
{ "name": string, "canonicalIngredientId": string | null, "estimatedGrams": number | null, "confidence": number }
],
"confidence": number
}
Språk: ${lang}.`;
}
private buildNutritionPrompt(locale: LocaleContext): string {
const lang = locale.languageTag.startsWith("en") ? "English" : "Swedish";
return `Du är en OCR-assistent för svenska livsmedelsförpackningar. Bilden visar en näringsdeklaration/ingredienslista.
Läs av exakt det som står på etiketten. Gissa aldrig värden som inte syns. Returnera:
- basis: "per_100_g", "per_100_ml" eller "per_portion" vad gäller tabellen?
- values: { kcal, proteinG, carbsG, fatG, saturatedFatG, fiberG, sugarG, saltG } null om ej synligt.
- ingredientsText: rå text med ingredienser, eller null.
- allergensDeclared: lista med allergener som explicit nämns (t.ex. ["gluten", "mjölk", "nötter"]).
- gtin: streckkod/EAN om synlig, annars null.
- productName: produktnamn om synligt, annars null.
- brand: varumärke om synligt, annars null.
- confidence: 0.01.0.
Svara ENDAST med giltig JSON som matchar detta schema:
{
"basis": "per_100_g" | "per_100_ml" | "per_portion" | null,
"values": { "kcal": number|null, "proteinG": number|null, "carbsG": number|null, "fatG": number|null, "saturatedFatG": number|null, "fiberG": number|null, "sugarG": number|null, "saltG": number|null } | null,
"ingredientsText": string | null,
"allergensDeclared": string[],
"gtin": string | null,
"productName": string | null,
"brand": string | null,
"confidence": number
}
Språk: ${lang}.`;
}
private buildExpiryPrompt(locale: LocaleContext): string {
const lang = locale.languageTag.startsWith("en") ? "English" : "Swedish";
return `Du är en OCR-assistent för svenska livsmedelsförpackningar. Bilden visar ett bäst-före-datum eller sista förbrukningsdatum.
Returnera:
- date: datum i format YYYY-MM-DD, eller null om ej läsbart.
- dateKind: "best_before" för "bäst före", "use_by" för "sista förbrukningsdag", annars null.
- confidence: 0.01.0.
Svara ENDAST med giltig JSON:
{
"date": "YYYY-MM-DD" | null,
"dateKind": "best_before" | "use_by" | null,
"confidence": number
}
Språk: ${lang}.`;
}
private buildReceiptPrompt(locale: LocaleContext): string {
const lang = locale.languageTag.startsWith("en") ? "English" : "Swedish";
return `Du är en OCR-assistent för svenska kvitton. Bilden visar ett butikskvitto.
Läs av:
- storeName: butikens namn, eller null.
- purchaseDate: datum i format YYYY-MM-DD, eller null.
- lines: varje rad med { rawText (exakt text), normalizedName (förenklat livsmedelsnamn eller null), canonicalIngredientId (alltid null), quantity (antal/enhet om går att utläsa, annars null), unit (enhetskod eller null), unitPriceMinor (ören per styck om synligt, annars null), totalPriceMinor (radens totalpris i ören om synligt, annars null), isDiscount (true om det är en rabatt/avdrag-rad), confidence (0.01.0) }.
- totalMinor: kvittots totalsumma i ören, eller null.
- discountTotalMinor: total rabatt i ören, eller null.
- confidence: 0.01.0 för hela tolkningen.
Svara ENDAST med giltig JSON som matchar detta schema:
{
"storeName": string | null,
"purchaseDate": "YYYY-MM-DD" | null,
"lines": [
{
"rawText": string,
"normalizedName": string | null,
"canonicalIngredientId": null,
"quantity": number | null,
"unit": string | null,
"unitPriceMinor": number | null,
"totalPriceMinor": number | null,
"isDiscount": boolean,
"confidence": number
}
],
"totalMinor": number | null,
"discountTotalMinor": number | null,
"confidence": number
}
Använd endast de enhetskoder du kan utläsa (GRAM, KILOGRAM, LITER, etc.), annars null. quantity är alltid ett decimaltal.
Språk: ${lang}.`;
}
+3 -3
View File
@@ -127,12 +127,12 @@ describe("GeminiAamosClient", () => {
});
const result = await client.runTask(
"READ_RECEIPT",
{ imageUrls: ["http://localhost/v1/mock-s3/receipt.jpg"], marketLocale: "sv-SE" },
"DEDUPLICATE_INVENTORY",
{ items: [] },
);
expect(result.status).toBe("failed");
expect(result.error).toContain("READ_RECEIPT");
expect(result.error).toContain("DEDUPLICATE_INVENTORY");
});
it("validates input before calling Gemini", async () => {