Add optional Claude system prompt cache.

7322dd1954fa20c08a0d370515407ea002bf3fea

Cohee <18619528+Cohee1207@users.noreply.github.com>

2 files changed, +27 -12Ignore whitespace
default/config.yaml+14 -5
@@ -41,6 +41,11 @@ enableCorsProxy: false
4141enableUserAccounts: false
4242# Enable discreet login mode: hides user list on the login screen
4343enableDiscreetLogin: false
44+# User session timeout *in seconds* (defaults to 24 hours).
45+## Set to a positive number to expire session after a certain time of inactivity
46+## Set to 0 to expire session when the browser is closed
47+## Set to a negative number to disable session expiration
48+sessionTimeout: 86400
4449# Used to sign session cookies. Will be auto-generated if not set
4550cookieSecret: ''
4651# Disable CSRF protection - NOT RECOMMENDED
@@ -123,10 +128,14 @@ ollama:
123128 # * 0: Unload the model immediately after the request
124129 # * 5m: Keep the model loaded for 5 minutes after the request. Accepts duration strings (e.g. 5h30m40s)
125130 keepAlive: -1
131+# -- ANTHROPIC CLAUDE API CONFIGURATION --
132+claude:
133+ # Enables caching of the system prompt (if supported).
134+ # https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
135+ # -- IMPORTANT! --
136+ # Use only when the prompt before the chat history is static and doesn't change between requests
137+ # (e.g {{random}} macro or lorebooks not as in-chat injections).
138+ # Otherwise, you'll just waste money on cache misses.
139+ enableSystemPromptCache: false
126140# -- SERVER PLUGIN CONFIGURATION --
127141enableServerPlugins: false
128-# User session timeout *in seconds* (defaults to 24 hours).
129-## Set to a positive number to expire session after a certain time of inactivity
130-## Set to 0 to expire session when the browser is closed
131-## Set to a negative number to disable session expiration
132-sessionTimeout: 86400
src/endpoints/backends/chat-completions.js+13 -7
@@ -105,6 +105,7 @@ async function sendClaudeRequest(request, response) {
105105 const apiUrl = new URL(request.body.reverse_proxy || API_CLAUDE).toString();
106106 const apiKey = request.body.reverse_proxy ? request.body.proxy_password : readSecret(request.user.directories, SECRET_KEYS.CLAUDE);
107107 const divider = '-'.repeat(process.stdout.columns);
108+ const enableSystemPromptCache = getConfigValue('claude.enableSystemPromptCache', false);
108109
109110 if (!apiKey) {
110111 console.log(color.red(`Claude API key is missing.\n${divider}`));
@@ -118,8 +119,8 @@ async function sendClaudeRequest(request, response) {
118119 controller.abort();
119120 });
120121 const additionalHeaders = {};
121122 letconst use_system_promptuseSystemPrompt = (request.body.model.startsWith('claude-2') || request.body.model.startsWith('claude-3')) && request.body.claude_use_sysprompt;
122123 letconst converted_promptconvertedPrompt = convertClaudeMessages(request.body.messages, request.body.assistant_prefill, use_system_promptuseSystemPrompt, request.body.human_sysprompt_message, request.body.char_name, request.body.user_name);
123124 // Add custom stop sequences
124125 const stopSequences = [];
125126 if (Array.isArray(request.body.stop)) {
@@ -127,7 +128,7 @@ async function sendClaudeRequest(request, response) {
127128 }
128129
129130 const requestBody = {
130131 messages: converted_promptconvertedPrompt.messages,
131132 model: request.body.model,
132133 max_tokens: request.body.max_tokens,
133134 stop_sequences: stopSequences,
@@ -136,13 +137,15 @@ async function sendClaudeRequest(request, response) {
136137 top_k: request.body.top_k,
137138 stream: request.body.stream,
138139 };
139140 if (use_system_promptuseSystemPrompt) {
140141 requestBody.system = converted_prompt.systemPrompt;enableSystemPromptCache
142+ ? [{ type: 'text', text: convertedPrompt.systemPrompt, cache_control: { type: 'ephemeral' } }]
143+ : convertedPrompt.systemPrompt;
141144 }
142145 if (Array.isArray(request.body.tools) && request.body.tools.length > 0) {
143146 // Claude doesn't do prefills on function calls, and doesn't allow empty messages
144147 if (converted_promptconvertedPrompt.messages.length && converted_promptconvertedPrompt.messages[converted_promptconvertedPrompt.messages.length - 1].role === 'assistant') {
145148 converted_promptconvertedPrompt.messages.push({ role: 'user', content: '.' });
146149 }
147150 additionalHeaders['anthropic-beta'] = 'tools-2024-05-16';
148151 requestBody.tool_choice = { type: request.body.tool_choice === 'required' ? 'any' : 'auto' };
@@ -151,6 +154,9 @@ async function sendClaudeRequest(request, response) {
151154 .map(tool => tool.function)
152155 .map(fn => ({ name: fn.name, description: fn.description, input_schema: fn.parameters }));
153156 }
157+ if (enableSystemPromptCache) {
158+ additionalHeaders['anthropic-beta'] = 'prompt-caching-2024-07-31';
159+ }
154160 console.log('Claude request:', requestBody);
155161
156162 const generateResponse = await fetch(apiUrl + '/messages', {