fix: add exponential backoff retry for LLM rate limits

apps/web/utils/llms/index.ts

@@ -35,7 +35,7 @@ import {
 } from "@/utils/error";
 import { getModel, type ModelType } from "@/utils/llms/model";
 import { createScopedLogger } from "@/utils/logger";
-import { withNetworkRetry } from "./retry";
+import { withNetworkRetry, withLLMRetry } from "./retry";
 
 const logger = createScopedLogger("llms");
 
@@ -97,9 +97,10 @@ export function createGenerateText({
     };
 
     try {
-      return await withNetworkRetry(() => generate(modelOptions.model), {
-        label,
-      });
+      return await withLLMRetry(
+        () => withNetworkRetry(() => generate(modelOptions.model), { label }),
+        { label },
+      );
     } catch (error) {
       // Try backup model for service unavailable or throttling errors
       if (
@@ -112,8 +113,11 @@ export function createGenerateText({
         });
 
         try {
-          return await withNetworkRetry(
-            () => generate(modelOptions.backupModel!),
+          return await withLLMRetry(
+            () =>
+              withNetworkRetry(() => generate(modelOptions.backupModel!), {
+                label,
+              }),
             { label },
           );
         } catch (backupError) {
@@ -202,13 +206,16 @@ export function createGenerateObject({
     };
 
     try {
-      return await withNetworkRetry(generate, {
-        label,
-        // Also retry on validation errors for generateObject
-        shouldRetry: (error) =>
-          NoObjectGeneratedError.isInstance(error) ||
-          TypeValidationError.isInstance(error),
-      });
+      return await withLLMRetry(
+        () =>
+          withNetworkRetry(generate, {
+            label,
+            shouldRetry: (error) =>
+              NoObjectGeneratedError.isInstance(error) ||
+              TypeValidationError.isInstance(error),
+          }),
+        { label },
+      );
     } catch (error) {
       await handleError(
         error,

apps/web/utils/llms/retry.test.ts

@@ -0,0 +1,268 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import { extractLLMErrorInfo, withLLMRetry } from "./retry";
+
+vi.mock("server-only", () => ({}));
+
+vi.mock("@/utils/sleep", () => ({
+  sleep: vi.fn().mockResolvedValue(undefined),
+}));
+
+function createError(
+  message: string,
+  props: { status?: number; code?: string } = {},
+): Error {
+  const error = new Error(message);
+  (error as unknown as { cause: typeof props }).cause = props;
+  return error;
+}
+
+describe("extractLLMErrorInfo", () => {
+  describe("rate limit detection", () => {
+    it("detects 429 status as rate limit", () => {
+      const error = { status: 429, message: "Too many requests" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects 429 in nested cause", () => {
+      const error = { cause: { status: 429, message: "Rate limited" } };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects OpenAI rate_limit_exceeded code", () => {
+      const error = { code: "rate_limit_exceeded", message: "Rate limit" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects Anthropic rate_limit_error code", () => {
+      const error = { code: "rate_limit_error", message: "Rate limit" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects Google RESOURCE_EXHAUSTED code", () => {
+      const error = { code: "RESOURCE_EXHAUSTED", message: "Quota exceeded" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects rate limit in error message", () => {
+      const error = { message: "You have hit a rate limit" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects quota exceeded in error message", () => {
+      const error = { message: "Quota exceeded for this API key" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects too many requests in error message", () => {
+      const error = { message: "Too many requests, please slow down" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(true);
+      expect(result.retryable).toBe(true);
+    });
+  });
+
+  describe("server error detection", () => {
+    it("detects 500 as server error", () => {
+      const error = { status: 500, message: "Internal server error" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.isRateLimit).toBe(false);
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects 502 as server error", () => {
+      const error = { status: 502, message: "Bad gateway" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects 503 as server error", () => {
+      const error = { status: 503, message: "Service unavailable" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects 504 as server error", () => {
+      const error = { status: 504, message: "Gateway timeout" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryable).toBe(true);
+    });
+
+    it("detects internal error in message", () => {
+      const error = { message: "An internal error occurred" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryable).toBe(true);
+    });
+  });
+
+  describe("non-retryable errors", () => {
+    it("does not retry 400 bad request", () => {
+      const error = { status: 400, message: "Invalid request" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryable).toBe(false);
+    });
+
+    it("does not retry 401 unauthorized", () => {
+      const error = { status: 401, message: "Invalid API key" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryable).toBe(false);
+    });
+
+    it("does not retry 403 forbidden", () => {
+      const error = { status: 403, message: "Access denied" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryable).toBe(false);
+    });
+  });
+
+  describe("retry-after extraction", () => {
+    it("extracts retry-after header in seconds", () => {
+      const error = {
+        status: 429,
+        cause: {
+          response: {
+            headers: { "retry-after": "30" },
+          },
+        },
+      };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryAfterMs).toBe(30_000);
+    });
+
+    it("extracts x-ratelimit-reset-requests header", () => {
+      const error = {
+        status: 429,
+        cause: {
+          response: {
+            headers: { "x-ratelimit-reset-requests": "60" },
+          },
+        },
+      };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryAfterMs).toBe(60_000);
+    });
+
+    it("extracts retry time from error message", () => {
+      const error = {
+        status: 429,
+        message: "Rate limited. Retry after 45 seconds",
+      };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryAfterMs).toBe(45_000);
+    });
+
+    it("returns undefined retryAfterMs when no retry info", () => {
+      const error = { status: 429, message: "Rate limited" };
+      const result = extractLLMErrorInfo(error);
+
+      expect(result.retryAfterMs).toBeUndefined();
+    });
+  });
+});
+
+describe("withLLMRetry", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("returns result on first successful attempt", async () => {
+    const fn = vi.fn().mockResolvedValue("success");
+
+    const result = await withLLMRetry(fn, { label: "test" });
+
+    expect(result).toBe("success");
+    expect(fn).toHaveBeenCalledTimes(1);
+  });
+
+  it("retries on rate limit error and succeeds", async () => {
+    const rateLimitError = createError("Rate limited", { status: 429 });
+    const fn = vi
+      .fn()
+      .mockRejectedValueOnce(rateLimitError)
+      .mockResolvedValueOnce("success after retry");
+
+    const result = await withLLMRetry(fn, { label: "test" });
+
+    expect(result).toBe("success after retry");
+    expect(fn).toHaveBeenCalledTimes(2);
+  });
+
+  it("retries on server error and succeeds", async () => {
+    const serverError = createError("Service unavailable", { status: 503 });
+    const fn = vi
+      .fn()
+      .mockRejectedValueOnce(serverError)
+      .mockResolvedValueOnce("success after retry");
+
+    const result = await withLLMRetry(fn, { label: "test" });
+
+    expect(result).toBe("success after retry");
+    expect(fn).toHaveBeenCalledTimes(2);
+  });
+
+  it("throws immediately on non-retryable errors", async () => {
+    const authError = createError("Invalid API key", { status: 401 });
+    const fn = vi.fn().mockRejectedValue(authError);
+
+    await expect(withLLMRetry(fn, { label: "test" })).rejects.toMatchObject({
+      error: { message: "Invalid API key" },
+    });
+    expect(fn).toHaveBeenCalledTimes(1);
+  });
+
+  it("uses custom maxRetries", async () => {
+    const rateLimitError = createError("Rate limited", { status: 429 });
+    const fn = vi.fn().mockRejectedValue(rateLimitError);
+
+    await expect(
+      withLLMRetry(fn, { label: "test", maxRetries: 1 }),
+    ).rejects.toThrow("Rate limited");
+
+    expect(fn).toHaveBeenCalledTimes(2);
+  });
+
+  it("calls sleep with delay on retry", async () => {
+    const { sleep } = await import("@/utils/sleep");
+    const rateLimitError = createError("Rate limited", { status: 429 });
+    const fn = vi
+      .fn()
+      .mockRejectedValueOnce(rateLimitError)
+      .mockResolvedValueOnce("success");
+
+    await withLLMRetry(fn, { label: "test" });
+
+    expect(sleep).toHaveBeenCalled();
+  });
+});