83ea6e5c · permissionBRICK/SillyTavern

public/index.html+8 -2 Viewed

                                             <div class="toggle-description justifyLeft marginBot5" data-source="claude" data-i18n="Allocates a portion of the response length for thinking (min: 1024 tokens, low: 10%, medium: 25%, high: 50%, max: 95%), but minimum 1024 tokens. Auto does not request thinking.">
                                                 Allocates a portion of the response length for thinking (min: 1024 tokens, low: 10%, medium: 25%, high: 50%, max: 95%), but minimum 1024 tokens. Auto does not request thinking.
                                             </div>
-                                            <div class="toggle-description justifyLeft marginBot5" data-source="makersuite,vertexai" data-i18n="Allocates a portion of the response length for thinking (Flash 2.5/Pro 2.5) (min: 0/128 tokens, low: 10%, medium: 25%, high: 50%, max: 24576/32768 tokens). Auto lets the model decide.">
+                                            <div class="toggle-description justifyLeft marginBot5" data-source="makersuite,vertexai">
-                                                Allocates a portion of the response length for thinking (Flash 2.5/Pro 2.5) (min: 0/128 tokens, low: 10%, medium: 25%, high: 50%, max: 24576/32768 tokens). Auto lets the model decide.
+                                                <span data-i18n="Sets a dynamic reasoning depth level for thinking (Flash 3/Pro 3). High and low are supported by both, minimal and medium are Flash 3 only. Auto lets the model decide.">
+                                                    Sets a dynamic reasoning depth level for thinking (Flash 3/Pro 3). High and low are supported by both, minimal and medium are Flash 3 only. Auto lets the model decide.
+                                                </span>
+                                                <br>
+                                                <span data-i18n="Allocates a portion of the response length for thinking (Flash 2.5/Pro 2.5) (min: 0/128 tokens, low: 10%, medium: 25%, high: 50%, max: 24576/32768 tokens). Auto lets the model decide.">
+                                                    Allocates a portion of the response length for thinking (Flash 2.5/Pro 2.5) (min: 0/128 tokens, low: 10%, medium: 25%, high: 50%, max: 24576/32768 tokens). Auto lets the model decide.
+                                                </span>
                                             </div>
                                         </div>
                                     </div>

src/endpoints/backends/chat-completions.js+5 -1 Viewed

             const thinkingConfig = { includeThoughts: includeReasoning };
             const thinkingBudget = calculateGoogleBudgetTokens(generationConfig.maxOutputTokens, reasoningEffort, model);
-            if (Number.isInteger(thinkingBudget)) {
+            if (typeof thinkingBudget === 'number' && Number.isInteger(thinkingBudget)) {
                 thinkingConfig.thinkingBudget = thinkingBudget;
+            }
+            if (typeof thinkingBudget === 'string' && thinkingBudget.length > 0) {
+                thinkingConfig.thinkingLevel = thinkingBudget;
+            }
             // Vertex doesn't allow mixing disabled thinking with includeThoughts
             if (useVertexAi && thinkingBudget === 0 && thinkingConfig.includeThoughts) {
                 console.info('Thinking budget is 0, but includeThoughts is true. Thoughts will not be included in the response.');

src/prompt-converters.js+50 -4 Viewed

  * @param {number} maxTokens Maximum tokens
  * @param {string} reasoningEffort Reasoning effort
  * @param {string} model Model name
- * @returns {number?} Budget tokens
+ * @returns {number|string|null} Budget tokens
  */
 export function calculateGoogleBudgetTokens(maxTokens, reasoningEffort, model) {
     function getFlashBudget() {

         return budgetTokens;
+    }
-    if (model.includes('flash-lite')) {
+    function getGemini3FlashBudget() {
+        switch (reasoningEffort) {
+            case REASONING_EFFORT.auto:
+                return null;
+            case REASONING_EFFORT.min:
+                return 'minimal';
+            case REASONING_EFFORT.low:
+                return 'low';
+            case REASONING_EFFORT.medium:
+                return 'medium';
+            case REASONING_EFFORT.high:
+                return 'high';
+            case REASONING_EFFORT.max:
+                return 'high';
+        }
+        return null;
+    }
+    function getGemini3ProBudget() {
+        switch (reasoningEffort) {
+            case REASONING_EFFORT.auto:
+                return null;
+            case REASONING_EFFORT.min:
+                return 'low';
+            case REASONING_EFFORT.low:
+                return 'low';
+            case REASONING_EFFORT.medium:
+                return 'low';
+            case REASONING_EFFORT.high:
+                return 'high';
+            case REASONING_EFFORT.max:
+                return 'high';
+        }
+        return null;
+    }
+    if (/gemini-3-pro/.test(model)) {
+        return getGemini3ProBudget();
+    }
+    if (/gemini-3-flash/.test(model) ) {
+        return getGemini3FlashBudget();
+    }
+    if (/flash-lite/.test(model)) {
         return getFlashLiteBudget();
+    }
-    if (model.includes('flash')) {
+    if (/flash/.test(model)) {
         return getFlashBudget();
+    }
-    if (model.includes('pro')) {
+    if (/pro/.test(model)) {
         return getProBudget();
+    }

Better thonk effort for Gem 3