Skip to content

Clean up anthropic caching #855

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 10, 2025
Merged

Clean up anthropic caching #855

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions app/lib/.server/llm/convex-agent.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,6 +133,16 @@ export async function convexAgent(args: {
};
}

if (modelProvider === 'Anthropic') {
messagesForDataStream[messagesForDataStream.length - 1].providerOptions = {
anthropic: {
cacheControl: {
type: 'ephemeral',
},
},
};
}

const dataStream = createDataStream({
execute(dataStream) {
const result = streamText({
Expand Down
69 changes: 6 additions & 63 deletions app/lib/.server/llm/provider.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,6 @@ import { getEnv } from '~/lib/.server/env';
// workaround for Vercel environment from
// https://github.com/vercel/ai/issues/199#issuecomment-1605245593
import { fetch } from '~/lib/.server/fetch';
import { GENERAL_SYSTEM_PROMPT_PRELUDE, ROLE_SYSTEM_PROMPT } from 'chef-agent/prompts/system';

const ALLOWED_AWS_REGIONS = ['us-east-1', 'us-east-2', 'us-west-2'];

Expand Down Expand Up @@ -161,8 +160,6 @@ export function getProvider(
// Falls back to the low Quality-of-Service Anthropic API key if the primary key is rate limited
const rateLimitAwareFetch = () => {
return async (input: RequestInfo | URL, init?: RequestInit) => {
const enrichedOptions = anthropicInjectCacheControl(init);

const throwIfBad = async (response: Response, isLowQos: boolean) => {
if (response.ok) {
return response;
Expand All @@ -181,7 +178,7 @@ export function getProvider(
throw new Error(JSON.stringify({ error: 'The model hit an error. Try sending your message again.' }));
};

const response = await fetch(input, enrichedOptions);
const response = await fetch(input, init);

if (response.status !== 429 && response.status !== 529) {
return throwIfBad(response, false);
Expand All @@ -201,12 +198,12 @@ export function getProvider(
response,
},
});
if (enrichedOptions && enrichedOptions.headers) {
const headers = new Headers(enrichedOptions.headers);
if (init && init.headers) {
const headers = new Headers(init.headers);
headers.set('x-api-key', lowQosKey);
enrichedOptions.headers = headers;
init.headers = headers;
}
const lowQosResponse = await fetch(input, enrichedOptions);
const lowQosResponse = await fetch(input, init);
return throwIfBad(lowQosResponse, true);
};
};
Expand All @@ -228,8 +225,7 @@ export function getProvider(

const userKeyApiFetch = (provider: ModelProvider) => {
return async (input: RequestInfo | URL, init?: RequestInit) => {
const requestInit = provider === 'Anthropic' ? anthropicInjectCacheControl(init) : init;
const result = await fetch(input, requestInit);
const result = await fetch(input, init);
if (result.status === 401) {
const text = await result.text();
throw new Error(JSON.stringify({ error: 'Invalid API key', details: text }));
Expand Down Expand Up @@ -273,56 +269,3 @@ const userKeyApiFetch = (provider: ModelProvider) => {
return result;
};
};

// sujayakar, 2025-03-25: This is mega-hax, but I can't figure out
// how to get the AI SDK to pass the cache control header to
// Anthropic with the `streamText` function. Setting
// `providerOptions.anthropic.cacheControl` doesn't seem to do
// anything. So, we instead directly inject the cache control
// header into the body of the request.
//
// tom, 2025-04-25: This is still an outstanding bug
// https://github.com/vercel/ai/issues/5942
function anthropicInjectCacheControl(options?: RequestInit) {
const start = Date.now();
if (!options) {
return options;
}
if (options.method !== 'POST') {
return options;
}
const headers = options.headers;
if (!headers) {
return options;
}
const contentType = new Headers(headers).get('content-type');
if (contentType !== 'application/json') {
return options;
}
if (typeof options.body !== 'string') {
throw new Error('Body must be a string');
}

const body = JSON.parse(options.body);

if (body.system[0].text !== ROLE_SYSTEM_PROMPT) {
throw new Error('First system message must be the role system prompt');
}
if (!body.system[1].text.startsWith(GENERAL_SYSTEM_PROMPT_PRELUDE)) {
throw new Error('Second system message must be the general system prompt prelude');
}
// Cache system prompt.
body.system[1].cache_control = { type: 'ephemeral' };
// Cache relevant files in system messages that are the same for all LLM requests after a user message.
body.system[body.system.length - 1].cache_control = { type: 'ephemeral' };

// Cache all messages.
const lastMessage = body.messages[body.messages.length - 1];
const lastMessagePartIndex = lastMessage.content.length - 1;
lastMessage.content[lastMessagePartIndex].cache_control = { type: 'ephemeral' };
body.messages[body.messages.length - 1].content[lastMessagePartIndex].cache_control = { type: 'ephemeral' };

const newBody = JSON.stringify(body);
console.log(`Injected system messages in ${Date.now() - start}ms`);
return { ...options, body: newBody };
}