Add support of stop strings to OpenAI / Claude

public/scripts/openai.js

@@ -34,6 +34,7 @@ import {
 } from "./PromptManager.js";
 
 import {
+    getCustomStoppingStrings,
     persona_description_positions,
     power_user,
 } from "./power-user.js";
@@ -120,6 +121,7 @@ const j2_max_topk = 10.0;
 const j2_max_freq = 5.0;
 const j2_max_pres = 5.0;
 const openrouter_website_model = 'OR_Website';
+const openai_max_stop_strings = 4;
 
 let biasCache = undefined;
 let model_list = [];
@@ -1138,6 +1140,7 @@ async function sendOpenAIRequest(type, openai_msgs_tosend, signal) {
         "max_tokens": oai_settings.openai_max_tokens,
         "stream": stream,
         "logit_bias": logit_bias,
+        "stop": getCustomStoppingStrings(openai_max_stop_strings),
     };
 
     // Proxy is only supported for Claude and OpenAI
@@ -1151,6 +1154,7 @@ async function sendOpenAIRequest(type, openai_msgs_tosend, signal) {
         generate_data['use_claude'] = true;
         generate_data['top_k'] = Number(oai_settings.top_k_openai);
         generate_data['exclude_assistant'] = oai_settings.exclude_assistant;
+        generate_data['stop'] = getCustomStoppingStrings(); // Claude shouldn't have limits on stop strings.
         // Don't add a prefill on quiet gens (summarization)
         if (!isQuiet && !oai_settings.exclude_assistant) {
             generate_data['assistant_prefill'] = substituteParams(oai_settings.assistant_prefill);

public/scripts/power-user.js

@@ -1537,8 +1537,19 @@ function setAvgBG() {
 
 }
 
-export function getCustomStoppingStrings() {
+
+/**
+ * Gets the custom stopping strings from the power user settings.
+ * @param {number | undefined} limit Number of strings to return. If undefined, returns all strings.
+ * @returns {string[]} An array of custom stopping strings
+ */
+export function getCustomStoppingStrings(limit = undefined) {
     try {
+        // If there's no custom stopping strings, return an empty array
+        if (!power_user.custom_stopping_strings) {
+            return [];
+        }
+
         // Parse the JSON string
         const strings = JSON.parse(power_user.custom_stopping_strings);
 
@@ -1547,8 +1558,8 @@ export function getCustomStoppingStrings() {
             return [];
         }
 
-        // Make sure all the elements are strings
+        // Make sure all the elements are strings. Apply the limit.
-        return strings.filter((s) => typeof s === 'string');
+        return strings.filter((s) => typeof s === 'string').slice(0, limit);
     } catch (error) {
         // If there's an error, return an empty array
         console.warn('Error parsing custom stopping strings:', error);

server.js

@@ -3391,6 +3391,12 @@ async function sendClaudeRequest(request, response) {
         }
 
         console.log('Claude request:', requestPrompt);
+        const stop_sequences = ["\n\nHuman:", "\n\nSystem:", "\n\nAssistant:"];
+
+        // Add custom stop sequences
+        if (Array.isArray(request.body.stop)) {
+            stop_sequences.push(...request.body.stop);
+        }
 
         const generateResponse = await fetch(api_url + '/complete', {
             method: "POST",
@@ -3399,7 +3405,7 @@ async function sendClaudeRequest(request, response) {
                 prompt: requestPrompt,
                 model: request.body.model,
                 max_tokens_to_sample: request.body.max_tokens,
-                stop_sequences: ["\n\nHuman:", "\n\nSystem:", "\n\nAssistant:"],
+                stop_sequences: stop_sequences,
                 temperature: request.body.temperature,
                 top_p: request.body.top_p,
                 top_k: request.body.top_k,
@@ -3489,6 +3495,11 @@ app.post("/generate_openai", jsonParser, function (request, response_generate_op
         return response_generate_openai.status(401).send({ error: true });
     }
 
+    // Add custom stop sequences
+    if (Array.isArray(request.body.stop)) {
+        bodyParams['stop'] = request.body.stop;
+    }
+
     const isTextCompletion = Boolean(request.body.model && (request.body.model.startsWith('text-') || request.body.model.startsWith('code-')));
     const textPrompt = isTextCompletion ? convertChatMLPrompt(request.body.messages) : '';
     const endpointUrl = isTextCompletion ? `${api_url}/completions` : `${api_url}/chat/completions`;