Teach local AI from user category corrections

- Add MerchantCorrection model: upsert by merchantName, Category enum
- Check corrections DB first in suggestCategoryForMerchant (source: "learned",
  no confirmation required); falls through to rules then Ollama if no match
- Inject recent corrections as few-shot examples in the Ollama prompt so the
  model improves even for merchants not yet explicitly corrected
- Add POST /categories/correct route to persist corrections
- Detect category override on form save (suggestedCategory !== chosen category)
  and silently fire a correction — no extra UX required
- Fix test isolation: beforeEach re-applies vi.fn() defaults after restoreAllMocks

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

src/app/categories/correct/route.ts

@@ -0,0 +1,25 @@
+import { Category } from "@prisma/client";
+import { NextResponse } from "next/server";
+import { z } from "zod";
+
+import { saveMerchantCorrection } from "@/lib/merchant-corrections";
+
+const correctionSchema = z.object({
+  merchantName: z.string().trim().min(1).max(80),
+  category: z.nativeEnum(Category, { message: "Choose a valid category." }),
+});
+
+export async function POST(request: Request) {
+  const payload = await request.json().catch(() => null);
+  const parsed = correctionSchema.safeParse(payload);
+
+  if (!parsed.success) {
+    return NextResponse.json(
+      { error: parsed.error.issues[0]?.message ?? "Invalid correction." },
+      { status: 400 },
+    );
+  }
+
+  await saveMerchantCorrection(parsed.data.merchantName, parsed.data.category);
+  return NextResponse.json({ ok: true });
+}

src/components/expense-workspace.tsx

@@ -10,7 +10,7 @@ type SuggestionResponse = {
   message: string;
   merchantName: string;
   requiresConfirmation: boolean;
-  source: "rule" | "model" | "unavailable";
+  source: "rule" | "model" | "unavailable" | "learned";
 };
 
 type ExpenseRecord = {
@@ -49,6 +49,7 @@ export function ExpenseWorkspace({ categoryOptions }: Props) {
   const [suggestionMessage, setSuggestionMessage] = useState<string | null>(null);
   const [needsSuggestionConfirmation, setNeedsSuggestionConfirmation] = useState(false);
   const [lastSuggestedMerchant, setLastSuggestedMerchant] = useState("");
+  const [suggestedCategory, setSuggestedCategory] = useState<CategoryValue | null>(null);
 
   useEffect(() => {
     async function loadExpenses() {
@@ -90,8 +91,8 @@ export function ExpenseWorkspace({ categoryOptions }: Props) {
     setSuggestionMessage(suggestion.message);
 
     if (suggestion.category) {
-      const suggestedCategory = suggestion.category;
-      setFormState((current) => ({ ...current, category: suggestedCategory }));
+      setFormState((current) => ({ ...current, category: suggestion.category! }));
+      setSuggestedCategory(suggestion.category);
     }
 
     setNeedsSuggestionConfirmation(suggestion.requiresConfirmation);
@@ -108,6 +109,7 @@ export function ExpenseWorkspace({ categoryOptions }: Props) {
     setSuggestionMessage(null);
     setNeedsSuggestionConfirmation(false);
     setLastSuggestedMerchant("");
+    setSuggestedCategory(null);
     setError(null);
   }
 
@@ -122,6 +124,7 @@ export function ExpenseWorkspace({ categoryOptions }: Props) {
     setSuggestionMessage(null);
     setNeedsSuggestionConfirmation(false);
     setLastSuggestedMerchant("");
+    setSuggestedCategory(null);
     setError(null);
   }
 
@@ -133,6 +136,20 @@ export function ExpenseWorkspace({ categoryOptions }: Props) {
       return;
     }
 
+    // If the AI (model or learned) suggested a category and the user changed it,
+    // silently record the correction so future suggestions improve.
+    if (
+      lastSuggestedMerchant &&
+      suggestedCategory !== null &&
+      formState.category !== suggestedCategory
+    ) {
+      void fetch("/categories/correct", {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ merchantName: lastSuggestedMerchant, category: formState.category }),
+      });
+    }
+
     setBusy(true);
     setError(null);
 
@@ -166,6 +183,7 @@ export function ExpenseWorkspace({ categoryOptions }: Props) {
     setSuggestionMessage(null);
     setNeedsSuggestionConfirmation(false);
     setLastSuggestedMerchant("");
+    setSuggestedCategory(null);
   }
 
   async function handleDelete(id: string) {

src/lib/category-suggestion.test.ts

@@ -1,5 +1,11 @@
-import { afterEach, describe, expect, it, vi } from "vitest";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 
+vi.mock("@/lib/merchant-corrections", () => ({
+  getMerchantCorrection: vi.fn(),
+  getRecentCorrections: vi.fn(),
+}));
+
+import { getMerchantCorrection, getRecentCorrections } from "@/lib/merchant-corrections";
 import { getMerchantRuleCategory, suggestCategoryForMerchant } from "@/lib/category-suggestion";
 
 describe("getMerchantRuleCategory", () => {
@@ -10,10 +16,26 @@ describe("getMerchantRuleCategory", () => {
 });
 
 describe("suggestCategoryForMerchant", () => {
+  beforeEach(() => {
+    vi.mocked(getMerchantCorrection).mockResolvedValue(null);
+    vi.mocked(getRecentCorrections).mockResolvedValue([]);
+  });
+
   afterEach(() => {
     vi.restoreAllMocks();
   });
 
+  it("returns learned category without confirmation when a correction exists", async () => {
+    const { getMerchantCorrection } = await import("@/lib/merchant-corrections");
+    vi.mocked(getMerchantCorrection).mockResolvedValueOnce({ merchantName: "Blue Tokai", category: "FOOD" });
+
+    const suggestion = await suggestCategoryForMerchant("Blue Tokai");
+
+    expect(suggestion.category).toBe("FOOD");
+    expect(suggestion.source).toBe("learned");
+    expect(suggestion.requiresConfirmation).toBe(false);
+  });
+
   it("uses the local model for unknown merchants", async () => {
     vi.spyOn(globalThis, "fetch").mockResolvedValue({
       ok: true,

src/lib/category-suggestion.ts

@@ -1,7 +1,8 @@
 import { CATEGORY_VALUES, type CategoryValue } from "@/lib/categories";
+import { getMerchantCorrection, getRecentCorrections } from "@/lib/merchant-corrections";
 import { generateOllamaJson } from "@/lib/ollama";
 
-type SuggestionSource = "rule" | "model" | "unavailable";
+type SuggestionSource = "rule" | "model" | "unavailable" | "learned";
 
 export type CategorySuggestion = {
   category: CategoryValue | null;
@@ -51,6 +52,24 @@ function parseSuggestedCategory(raw: unknown): CategoryValue | null {
   return CATEGORY_VALUES.includes(normalized as CategoryValue) ? (normalized as CategoryValue) : null;
 }
 
+function buildOllamaPrompt(merchantName: string, examples: Array<{ merchantName: string; category: string }>) {
+  const lines = [
+    "You categorize personal expense merchants.",
+    "Return JSON with one key named category.",
+    "Allowed values only: RENT, FOOD, TRANSPORT, BILLS, SHOPPING, HEALTH, ENTERTAINMENT, MISC.",
+  ];
+
+  if (examples.length > 0) {
+    lines.push("Use these corrections the user has made previously as guidance:");
+    for (const ex of examples) {
+      lines.push(`  "${ex.merchantName}" → ${ex.category}`);
+    }
+  }
+
+  lines.push(`Merchant: ${merchantName}`);
+  return lines.join("\n");
+}
+
 export async function suggestCategoryForMerchant(merchantName: string): Promise<CategorySuggestion> {
   const normalized = normalizeMerchantName(merchantName);
 
@@ -64,6 +83,19 @@ export async function suggestCategoryForMerchant(merchantName: string): Promise<
     };
   }
 
+  // 1. Check stored user corrections first — highest priority, no confirmation needed.
+  const learned = await getMerchantCorrection(normalized);
+  if (learned) {
+    return {
+      category: learned.category as CategoryValue,
+      message: "Category auto-filled from your previous correction.",
+      merchantName: normalized,
+      requiresConfirmation: false,
+      source: "learned",
+    };
+  }
+
+  // 2. Hardcoded rules for well-known merchants.
   const matchedCategory = getMerchantRuleCategory(normalized);
   if (matchedCategory) {
     return {
@@ -75,11 +107,11 @@ export async function suggestCategoryForMerchant(merchantName: string): Promise<
     };
   }
 
+  // 3. Ask Ollama, providing recent user corrections as few-shot examples.
   try {
+    const recentCorrections = await getRecentCorrections(20);
     const parsed = await generateOllamaJson<{ category?: string }>({
-      prompt:
-        "You categorize personal expense merchants. Return JSON with one key named category. Allowed values only: RENT, FOOD, TRANSPORT, BILLS, SHOPPING, HEALTH, ENTERTAINMENT, MISC. Merchant: " +
-        normalized,
+      prompt: buildOllamaPrompt(normalized, recentCorrections),
     });
     const category = parseSuggestedCategory(parsed?.category);
 

src/lib/merchant-corrections.ts

@@ -0,0 +1,28 @@
+import type { Category } from "@prisma/client";
+
+import { db } from "@/lib/db";
+
+export type MerchantCorrection = {
+  merchantName: string;
+  category: Category;
+};
+
+export async function getMerchantCorrection(merchantName: string): Promise<MerchantCorrection | null> {
+  return db.merchantCorrection.findUnique({ where: { merchantName } });
+}
+
+export async function getRecentCorrections(limit = 30): Promise<MerchantCorrection[]> {
+  return db.merchantCorrection.findMany({
+    orderBy: { updatedAt: "desc" },
+    take: limit,
+    select: { merchantName: true, category: true },
+  });
+}
+
+export async function saveMerchantCorrection(merchantName: string, category: Category): Promise<void> {
+  await db.merchantCorrection.upsert({
+    where: { merchantName },
+    update: { category },
+    create: { merchantName, category },
+  });
+}