implement messages api and selector for nuclaude models

public/index.html

@@ -2348,6 +2348,8 @@
                                 <h4 data-i18n="Claude Model">Claude Model</h4>
                                 <select id="model_claude_select">
                                     <optgroup label="Latest">
+                                        <option value="claude-3-opus">claude-3-opus</option>
+                                        <option value="claude-3-sonnet">claude-3-sonnet</option>
                                         <option value="claude-2">claude-2</option>
                                         <option value="claude-v1">claude-v1</option>
                                         <option value="claude-v1-100k">claude-v1-100k</option>
@@ -2355,6 +2357,8 @@
                                         <option value="claude-instant-v1-100k">claude-instant-v1-100k</option>
                                     </optgroup>
                                     <optgroup label="Sub-versions">
+                                        <option value="claude-3-opus-20240229">claude-3-opus-20240229</option>
+                                        <option value="claude-3-sonnet-20240229">claude-3-sonnet-20240229</option>
                                         <option value="claude-2.1">claude-2.1</option>
                                         <option value="claude-2.0">claude-2.0</option>
                                         <option value="claude-v1.3">claude-v1.3</option>

public/scripts/openai.js

@@ -1751,7 +1751,7 @@ async function sendOpenAIRequest(type, messages, signal) {
 
 function getStreamingReply(data) {
     if (oai_settings.chat_completion_source == chat_completion_sources.CLAUDE) {
-        return data?.completion || '';
+        return data?.delta?.text || '';
     } else if (oai_settings.chat_completion_source == chat_completion_sources.MAKERSUITE) {
         return data?.candidates?.[0]?.content?.parts?.[0]?.text || '';
     } else {
@@ -3439,7 +3439,7 @@ async function onModelChange() {
         if (oai_settings.max_context_unlocked) {
             $('#openai_max_context').attr('max', max_200k);
         }
-        else if (value == 'claude-2.1' || value == 'claude-2') {
+        else if (value == 'claude-2.1' || value == 'claude-2' || value.startsWith('claude-3')) {
             $('#openai_max_context').attr('max', max_200k);
         }
         else if (value.endsWith('100k') || value.startsWith('claude-2') || value === 'claude-instant-1.2') {

src/endpoints/backends/chat-completions.js

@@ -5,7 +5,7 @@ const { Readable } = require('stream');
 const { jsonParser } = require('../../express-common');
 const { CHAT_COMPLETION_SOURCES, GEMINI_SAFETY, BISON_SAFETY, OPENROUTER_HEADERS } = require('../../constants');
 const { forwardFetchResponse, getConfigValue, tryParse, uuidv4, mergeObjectWithYaml, excludeKeysByYaml, color } = require('../../util');
-const { convertClaudePrompt, convertGooglePrompt, convertTextCompletionPrompt } = require('../prompt-converters');
+const { convertClaudeMessages, convertGooglePrompt, convertTextCompletionPrompt } = require('../prompt-converters');
 
 const { readSecret, SECRET_KEYS } = require('../secrets');
 const { getTokenizerModel, getSentencepiceTokenizer, getTiktokenTokenizer, sentencepieceTokenizers, TEXT_COMPLETION_MODELS } = require('../tokenizers');
@@ -34,45 +34,8 @@ async function sendClaudeRequest(request, response) {
         request.socket.on('close', function () {
             controller.abort();
         });
-
+        let use_system_prompt = (request.body.model.startsWith('claude-2') || request.body.model.startsWith('claude-3')) && request.body.claude_use_sysprompt;
-        const isSysPromptSupported = request.body.model === 'claude-2' || request.body.model === 'claude-2.1';
+        let converted_prompt = convertClaudeMessages(request.body.messages, request.body.assistant_prefill, use_system_prompt, request.body.human_sysprompt_message);
-        const requestPrompt = convertClaudePrompt(request.body.messages, !request.body.exclude_assistant, request.body.assistant_prefill, isSysPromptSupported, request.body.claude_use_sysprompt, request.body.human_sysprompt_message, request.body.claude_exclude_prefixes);
-
-        // Check Claude messages sequence and prefixes presence.
-        let sequenceError = [];
-        const sequence = requestPrompt.split('\n').filter(x => x.startsWith('Human:') || x.startsWith('Assistant:'));
-        const humanFound = sequence.some(line => line.startsWith('Human:'));
-        const assistantFound = sequence.some(line => line.startsWith('Assistant:'));
-        let humanErrorCount = 0;
-        let assistantErrorCount = 0;
-
-        for (let i = 0; i < sequence.length - 1; i++) {
-            if (sequence[i].startsWith(sequence[i + 1].split(':')[0])) {
-                if (sequence[i].startsWith('Human:')) {
-                    humanErrorCount++;
-                } else if (sequence[i].startsWith('Assistant:')) {
-                    assistantErrorCount++;
-                }
-            }
-        }
-
-        if (!humanFound) {
-            sequenceError.push(`${divider}\nWarning: No 'Human:' prefix found in the prompt.\n${divider}`);
-        }
-        if (!assistantFound) {
-            sequenceError.push(`${divider}\nWarning: No 'Assistant: ' prefix found in the prompt.\n${divider}`);
-        }
-        if (sequence[0] && !sequence[0].startsWith('Human:')) {
-            sequenceError.push(`${divider}\nWarning: The messages sequence should start with 'Human:' prefix.\nMake sure you have '\\n\\nHuman:' prefix at the very beggining of the prompt, or after the system prompt.\n${divider}`);
-        }
-        if (humanErrorCount > 0 || assistantErrorCount > 0) {
-            sequenceError.push(`${divider}\nWarning: Detected incorrect Prefix sequence(s).`);
-            sequenceError.push(`Incorrect "Human:" prefix(es): ${humanErrorCount}.\nIncorrect "Assistant: " prefix(es): ${assistantErrorCount}.`);
-            sequenceError.push('Check the prompt above and fix it in the SillyTavern.');
-            sequenceError.push('\nThe correct sequence in the console should look like this:\n(System prompt msg) <-(for the sysprompt format only, else have \\n\\n above the first human\'s  message.)');
-            sequenceError.push(`\\n +      <-----(Each message beginning with the "Assistant:/Human:" prefix must have \\n\\n before it.)\n\\n +\nHuman: \\n +\n\\n +\nAssistant: \\n +\n...\n\\n +\nHuman: \\n +\n\\n +\nAssistant: \n${divider}`);
-        }
-
         // Add custom stop sequences
         const stopSequences = ['\n\nHuman:', '\n\nSystem:', '\n\nAssistant:'];
         if (Array.isArray(request.body.stop)) {
@@ -80,23 +43,21 @@ async function sendClaudeRequest(request, response) {
         }
 
         const requestBody = {
-            prompt: requestPrompt,
+            messages: converted_prompt.messages,
             model: request.body.model,
-            max_tokens_to_sample: request.body.max_tokens,
+            max_tokens: request.body.max_tokens,
             stop_sequences: stopSequences,
             temperature: request.body.temperature,
             top_p: request.body.top_p,
             top_k: request.body.top_k,
             stream: request.body.stream,
         };
-
+        if (use_system_prompt) {
+            requestBody.system = converted_prompt.systemPrompt;
+        }
         console.log('Claude request:', requestBody);
 
-        sequenceError.forEach(sequenceError => {
+        const generateResponse = await fetch(apiUrl + '/messages', {
-            console.log(color.red(sequenceError));
-        });
-
-        const generateResponse = await fetch(apiUrl + '/complete', {
             method: 'POST',
             signal: controller.signal,
             body: JSON.stringify(requestBody),
@@ -118,7 +79,7 @@ async function sendClaudeRequest(request, response) {
             }
 
             const generateResponseJson = await generateResponse.json();
-            const responseText = generateResponseJson.completion;
+            const responseText = generateResponseJson.content[0].text;
             console.log('Claude response:', generateResponseJson);
 
             // Wrap it back to OAI format

src/endpoints/prompt-converters.js

@@ -74,48 +74,58 @@ function convertClaudePrompt(messages, addAssistantPostfix, addAssistantPrefill,
 /**
  * Convert ChatML objects into working with Anthropic's new Messaging API.
  * @param {object[]} messages Array of messages
- * @param {boolean}  addAssistantPostfix Add Assistant postfix.
+ * @param {string}   prefillString User determined prefill string
- * @param {string}   addAssistantPrefill Add Assistant prefill after the assistant postfix.
+ * @param {boolean}  useSysPrompt See if we want to use a system prompt
- * @param {boolean}  withSysPromptSupport Indicates if the Claude model supports the system prompt format.
+ * @param {string}   humanMsgFix Add Human message between system prompt and assistant.
- * @param {string}   addSysHumanMsg Add Human message between system prompt and assistant.
  */
-function convertClaudeMessages(messages, addAssistantPostfix, addAssistantPrefill, addSysHumanMsg) {
+function convertClaudeMessages(messages, prefillString, useSysPrompt, humanMsgFix) {
-    // Collect all the system messages up until the first instance of a non-system message, and then remove them from the messages array.
+    // Since the messaging endpoint only supports user assistant roles in turns, we have to merge messages with the same role if they follow eachother
-    let systemPrompt = '';
+    let mergedMessages = [];
-    let i;
-    for (i = 0; i < messages.length; i++) {
-        if (messages[i].role !== 'system') {
-            break;
-        }
-        systemPrompt += `${messages[i].content}\n\n`;
-    }
-
-    messages.splice(0, i);
-
-    // Check if the first message in the array is of type user, if not, interject with addSysHumanMsg or a blank message.
-    if (messages.length > 0 && messages[0].role !== 'user') {
-        messages.unshift({
-            role: 'user',
-            content: addSysHumanMsg || '',
-        });
-    }
-
-    // Now replace all further messages that have the role 'system' with the role 'user'.
     messages.forEach((message) => {
+        if (mergedMessages.length > 0 && mergedMessages[mergedMessages.length - 1].role === message.role) {
+            mergedMessages[mergedMessages.length - 1].content += '\n\n' + message.content;
+        } else {
+            mergedMessages.push(message);
+        }
+    });
+
+    let systemPrompt = '';
+    if (useSysPrompt) {
+        // Collect all the system messages up until the first instance of a non-system message, and then remove them from the messages array.
+        let i;
+        for (i = 0; i < mergedMessages.length; i++) {
+            if (mergedMessages[i].role !== 'system') {
+                break;
+            }
+            systemPrompt += `${mergedMessages[i].content}\n\n`;
+        }
+
+        mergedMessages.splice(0, i);
+
+        // Check if the first message in the array is of type user, if not, interject with humanMsgFix or a blank message.
+        if (mergedMessages.length > 0 && mergedMessages[0].role !== 'user') {
+            mergedMessages.unshift({
+                role: 'user',
+                content: humanMsgFix || '',
+            });
+        }
+    }
+    // Now replace all further messages that have the role 'system' with the role 'user'. (or all if we're not using one)
+    mergedMessages.forEach((message) => {
         if (message.role === 'system') {
             message.role = 'user';
         }
     });
 
-    // Postfix and prefill
+    // Shouldn't be conditional anymore, messages api expects the last role to be user unless we're explicitly prefilling
-    if (addAssistantPostfix) {
+    if (prefillString) {
-        messages.push({
+        mergedMessages.push({
             role: 'assistant',
-            content: addAssistantPrefill || '',
+            content: prefillString,
         });
     }
 
-    return { messages: messages, systemPrompt: systemPrompt.trim() };
+    return { messages: mergedMessages, systemPrompt: systemPrompt.trim() };
 }
 
 /**

src/endpoints/tokenizers.js

@@ -249,6 +249,7 @@ async function loadClaudeTokenizer(modelPath) {
 }
 
 function countClaudeTokens(tokenizer, messages) {
+    // Should be fine if we use the old conversion method instead of the messages API one i think?
     const convertedPrompt = convertClaudePrompt(messages, false, false, false);
 
     // Fallback to strlen estimation