Fallback to Haiku if service unavailable

apps/web/env.ts

@@ -49,6 +49,7 @@ export const env = createEnv({
     WEBHOOK_URL: z.string().optional(),
     INTERNAL_API_KEY: z.string().optional(),
     WHITELIST_FROM: z.string().optional(),
+    USE_BACKUP_MODEL: z.coerce.boolean().optional().default(false),
 
     // license
     LICENSE_1_SEAT_VARIANT_ID: z.coerce.number().optional(),
@@ -111,7 +112,10 @@ export const env = createEnv({
     NEXT_PUBLIC_AXIOM_TOKEN: z.string().optional(),
     NEXT_PUBLIC_BEDROCK_SONNET_MODEL: z
       .string()
-      .default("anthropic.claude-3-5-sonnet-20241022-v2:0"),
+      .default("us.anthropic.claude-3-5-sonnet-20241022-v2:0"),
+    NEXT_PUBLIC_BEDROCK_HAIKU_MODEL: z
+      .string()
+      .default("us.anthropic.claude-3-5-haiku-20241022-v1:0"),
     NEXT_PUBLIC_OLLAMA_MODEL: z.string().optional(),
   },
   // For Next.js >= 13.4.4, you only need to destructure client variables:
@@ -181,6 +185,8 @@ export const env = createEnv({
     NEXT_PUBLIC_AXIOM_TOKEN: process.env.NEXT_PUBLIC_AXIOM_TOKEN,
     NEXT_PUBLIC_BEDROCK_SONNET_MODEL:
       process.env.NEXT_PUBLIC_BEDROCK_SONNET_MODEL,
+    NEXT_PUBLIC_BEDROCK_HAIKU_MODEL:
+      process.env.NEXT_PUBLIC_BEDROCK_HAIKU_MODEL,
     NEXT_PUBLIC_OLLAMA_MODEL: process.env.NEXT_PUBLIC_OLLAMA_MODEL,
   },
 });

apps/web/utils/error.ts

@@ -108,6 +108,10 @@ export function isAWSThrottlingError(error: unknown): error is Error {
   );
 }
 
+export function isServiceUnavailableError(error: unknown): error is Error {
+  return error instanceof Error && error.name === "ServiceUnavailableException";
+}
+
 // we don't want to capture these errors in Sentry
 export function isKnownApiError(error: unknown): boolean {
   return (

apps/web/utils/llms/config.ts

@@ -12,6 +12,7 @@ export const Model = {
   GPT_4O: "gpt-4o",
   GPT_4O_MINI: "gpt-4o-mini",
   CLAUDE_3_5_SONNET_BEDROCK: env.NEXT_PUBLIC_BEDROCK_SONNET_MODEL,
+  CLAUDE_3_5_HAIKU_BEDROCK: env.NEXT_PUBLIC_BEDROCK_HAIKU_MODEL,
   CLAUDE_3_5_SONNET_ANTHROPIC: "claude-3-5-sonnet-20241022",
   ...(supportsOllama ? { OLLAMA: env.NEXT_PUBLIC_OLLAMA_MODEL } : {}),
 };

apps/web/utils/llms/index.ts

@@ -22,6 +22,7 @@ import {
   isInvalidOpenAIModelError,
   isOpenAIAPIKeyDeactivatedError,
   isOpenAIRetryError,
+  isServiceUnavailableError,
 } from "@/utils/error";
 import { sleep } from "@/utils/sleep";
 
@@ -81,21 +82,29 @@ function getModel({ aiProvider, aiModel, aiApiKey }: UserAIFields) {
   throw new Error("AI provider not supported");
 }
 
-export async function chatCompletionObject<T>({
-  userAi,
-  prompt,
-  system,
-  schema,
-  userEmail,
-  usageLabel,
-}: {
+type ChatCompletionObjectArgs<T> = {
   userAi: UserAIFields;
   prompt: string;
   system?: string;
   schema: z.Schema<T>;
   userEmail: string;
   usageLabel: string;
-}) {
+};
+
+export async function chatCompletionObject<T>(
+  options: ChatCompletionObjectArgs<T>,
+) {
+  return withBackupModel(chatCompletionObjectInternal, options);
+}
+
+async function chatCompletionObjectInternal<T>({
+  userAi,
+  prompt,
+  system,
+  schema,
+  userEmail,
+  usageLabel,
+}: ChatCompletionObjectArgs<T>) {
   try {
     const { provider, model, llmModel } = getModel(userAi);
 
@@ -141,7 +150,7 @@ export async function chatCompletionStream({
 }) {
   const { provider, model, llmModel } = getModel(userAi);
 
-  const result = await streamText({
+  const result = streamText({
     model: llmModel,
     prompt,
     system,
@@ -162,23 +171,29 @@ export async function chatCompletionStream({
   return result;
 }
 
-export async function chatCompletionTools({
-  userAi,
-  prompt,
-  system,
-  tools,
-  maxSteps,
-  label,
-  userEmail,
-}: {
+type ChatCompletionToolsArgs = {
   userAi: UserAIFields;
   prompt: string;
   system?: string;
   tools: Record<string, CoreTool>;
   maxSteps?: number;
   label: string;
   userEmail: string;
-}) {
+};
+
+export async function chatCompletionTools<T>(options: ChatCompletionToolsArgs) {
+  return withBackupModel(chatCompletionToolsInternal, options);
+}
+
+async function chatCompletionToolsInternal({
+  userAi,
+  prompt,
+  system,
+  tools,
+  maxSteps,
+  label,
+  userEmail,
+}: ChatCompletionToolsArgs) {
   try {
     const { provider, model, llmModel } = getModel(userAi);
 
@@ -292,6 +307,28 @@ export async function withRetry<T>(
   throw lastError;
 }
 
+// Helps when service is unavailable / throttled / rate limited
+async function withBackupModel<T, Args extends { userAi: UserAIFields }>(
+  fn: (args: Args) => Promise<T>,
+  args: Args,
+): Promise<T> {
+  try {
+    return await fn(args);
+  } catch (error) {
+    if (env.USE_BACKUP_MODEL && isServiceUnavailableError(error)) {
+      return await fn({
+        ...args,
+        userAi: {
+          aiProvider: Provider.ANTHROPIC,
+          aiModel: env.NEXT_PUBLIC_BEDROCK_HAIKU_MODEL,
+          aiApiKey: args.userAi.aiApiKey,
+        },
+      });
+    }
+    throw error;
+  }
+}
+
 async function handleError(error: unknown, userEmail: string) {
   if (APICallError.isInstance(error)) {
     if (isIncorrectOpenAIAPIKeyError(error)) {

apps/web/utils/usage.ts

@@ -83,6 +83,14 @@ const costs: Record<
     input: 3 / 1_000_000,
     output: 15 / 1_000_000,
   },
+  "anthropic.claude-3-5-haiku-20241022-v1:0": {
+    input: 0.8 / 1_000_000,
+    output: 4 / 1_000_000,
+  },
+  "us.anthropic.claude-3-5-haiku-20241022-v1:0": {
+    input: 0.8 / 1_000_000,
+    output: 4 / 1_000_000,
+  },
 };
 
 // returns cost in cents