rixty/SillyTavern
1
0
Fork 0
mirror of https://github.com/SillyTavern/SillyTavern.git synced 2025-06-05 21:59:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'staging' into vision-cleanup

Browse Source
This commit is contained in:
equal-l2
2025-04-25 01:49:40 +09:00
parent 903839c9c5 c16be2ec0e
commit 3fd12b28dc
5 changed files with 142 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
public/index.html
View File

@ -2048,16 +2048,20 @@
</span>
</div>
</div>
<div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai">
<div class="flex-container oneline-dropdown" title="Constrains effort on reasoning for reasoning models.&#10;Currently supported values are low, medium, and high.&#10;Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response." data-i18n="[title]Constrains effort on reasoning for reasoning models.">
<div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai,makersuite,openrouter">
<div class="flex-container oneline-dropdown" title="Constrains effort on reasoning for reasoning models.&#10;Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response." data-i18n="[title]Constrains effort on reasoning for reasoning models.">
<label for="openai_reasoning_effort">
<span data-i18n="Reasoning Effort">Reasoning Effort</span>
<i data-source="claude" class="opacity50p fa-solid fa-circle-info" title="Allocates a portion of the response length for thinking (low: 10%, medium: 25%, high: 50%), but minimum 1024 tokens."></i>
<i data-source="openai,custom,xai,openrouter" class="opacity50p fa-solid fa-circle-info" title="OpenAI-style options: low, medium, high. Minimum and maximum are aliased to low and high. Auto does not send an effort level."></i>
<i data-source="claude,makersuite" class="opacity50p fa-solid fa-circle-info" title="Allocates a portion of the response length for thinking (low: 10%, medium: 25%, high: 50%). Other options are model-dependent."></i>
</label>
<select id="openai_reasoning_effort">
<option data-i18n="openai_reasoning_effort_auto" value="auto">Auto</option>
<option data-i18n="openai_reasoning_effort_minimum" value="min">Mininum</option>
<option data-i18n="openai_reasoning_effort_low" value="low">Low</option>
<option data-i18n="openai_reasoning_effort_medium" value="medium">Medium</option>
<option data-i18n="openai_reasoning_effort_high" value="high">High</option>
<option data-i18n="openai_reasoning_effort_maximum" value="max">Maximum</option>
</select>
</div>
</div>

22
public/script.js
View File

@ -6955,15 +6955,23 @@ export async function saveChat({ chatName, withMetadata, mesId, force = false }
throw new Error(result.statusText);
}
const forceSaveConfirmed = await Popup.show.confirm(
t`ERROR: Chat integrity check failed.`,
t`Continuing the operation may result in data loss. Would you like to overwrite the chat file anyway? Pressing "NO" will cancel the save operation.`,
{ okButton: t`Yes, overwrite`, cancelButton: t`No, cancel` },
) === POPUP_RESULT.AFFIRMATIVE;
const popupResult = await Popup.show.input(
t`ERROR: Chat integrity check failed while saving the file.`,
t`<p>After you click OK, the page will be reloaded to prevent data corruption.</p>
<p>To confirm an overwrite (and potentially <b>LOSE YOUR DATA</b>), enter <code>OVERWRITE</code> (in all caps) in the box below before clicking OK.</p>`,
'',
{ okButton: 'OK', cancelButton: false },
);
if (forceSaveConfirmed) {
await saveChat({ chatName, withMetadata, mesId, force: true });
const forceSaveConfirmed = popupResult === 'OVERWRITE';
if (!forceSaveConfirmed) {
console.warn('Chat integrity check failed, and user did not confirm the overwrite. Reloading the page.');
window.location.reload();
return;
}
await saveChat({ chatName, withMetadata, mesId, force: true });
} catch (error) {
console.error(error);
toastr.error(t`Check the server connection and reload the page to prevent data loss.`, t`Chat could not be saved`);

40
public/scripts/openai.js
View File

@ -216,6 +216,15 @@ const openrouter_middleout_types = {
OFF: 'off',
};
export const reasoning_effort_types = {
auto: 'auto',
low: 'low',
medium: 'medium',
high: 'high',
min: 'min',
max: 'max',
};
const sensitiveFields = [
'reverse_proxy',
'proxy_password',
@ -382,7 +391,7 @@ const default_settings = {
continue_postfix: continue_postfix_types.SPACE,
custom_prompt_post_processing: custom_prompt_post_processing_types.NONE,
show_thoughts: true,
reasoning_effort: 'medium',
reasoning_effort: reasoning_effort_types.auto,
enable_web_search: false,
request_images: false,
seed: -1,
@ -463,7 +472,7 @@ const oai_settings = {
continue_postfix: continue_postfix_types.SPACE,
custom_prompt_post_processing: custom_prompt_post_processing_types.NONE,
show_thoughts: true,
reasoning_effort: 'medium',
reasoning_effort: reasoning_effort_types.auto,
enable_web_search: false,
request_images: false,
seed: -1,
@ -1937,6 +1946,31 @@ async function sendAltScaleRequest(messages, logit_bias, signal, type) {
return data.output;
}
function getReasoningEffort() {
// These sources expect the effort as string.
const reasoningEffortSources = [
chat_completion_sources.OPENAI,
chat_completion_sources.CUSTOM,
chat_completion_sources.XAI,
chat_completion_sources.OPENROUTER,
];
if (!reasoningEffortSources.includes(oai_settings.chat_completion_source)) {
return oai_settings.reasoning_effort;
}
switch (oai_settings.reasoning_effort) {
case reasoning_effort_types.auto:
return undefined;
case reasoning_effort_types.min:
return reasoning_effort_types.low;
case reasoning_effort_types.max:
return reasoning_effort_types.high;
default:
return oai_settings.reasoning_effort;
}
}
/**
* Send a chat completion request to backend
* @param {string} type (impersonate, quiet, continue, etc)
@ -2023,7 +2057,7 @@ async function sendOpenAIRequest(type, messages, signal) {
'char_name': name2,
'group_names': getGroupNames(),
'include_reasoning': Boolean(oai_settings.show_thoughts),
'reasoning_effort': String(oai_settings.reasoning_effort),
'reasoning_effort': getReasoningEffort(),
'enable_web_search': Boolean(oai_settings.enable_web_search),
'request_images': Boolean(oai_settings.request_images),
'custom_prompt_post_processing': oai_settings.custom_prompt_post_processing,

29
src/endpoints/backends/chat-completions.js
View File

@ -28,7 +28,8 @@ import {
cachingAtDepthForOpenRouterClaude,
cachingAtDepthForClaude,
getPromptNames,
calculateBudgetTokens,
calculateClaudeBudgetTokens,
calculateGoogleBudgetTokens,
} from '../../prompt-converters.js';
import { readSecret, SECRET_KEYS } from '../secrets.js';
@ -202,7 +203,7 @@ async function sendClaudeRequest(request, response) {
// No prefill when thinking
voidPrefill = true;
const reasoningEffort = request.body.reasoning_effort;
const budgetTokens = calculateBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream);
const budgetTokens = calculateClaudeBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream);
const minThinkTokens = 1024;
if (requestBody.max_tokens <= minThinkTokens) {
const newValue = requestBody.max_tokens + minThinkTokens;
@ -340,6 +341,7 @@ async function sendMakerSuiteRequest(request, response) {
const stream = Boolean(request.body.stream);
const enableWebSearch = Boolean(request.body.enable_web_search);
const requestImages = Boolean(request.body.request_images);
const reasoningEffort = String(request.body.reasoning_effort);
const isThinking = model.includes('thinking');
const isGemma = model.includes('gemma');
@ -412,13 +414,26 @@ async function sendMakerSuiteRequest(request, response) {
tools.push({ function_declarations: functionDeclarations });
}
// One more models list to maintain, yay
const thinkingBudgetModels = [
'gemini-2.5-flash-preview-04-17',
];
if (thinkingBudgetModels.includes(model)) {
const thinkingBudget = calculateGoogleBudgetTokens(generationConfig.maxOutputTokens, reasoningEffort);
if (Number.isInteger(thinkingBudget)) {
generationConfig.thinkingConfig = { thinkingBudget: thinkingBudget };
}
}
let body = {
contents: prompt.contents,
safetySettings: safetySettings,
generationConfig: generationConfig,
};
if (useSystemPrompt) {
if (useSystemPrompt && Array.isArray(prompt.system_instruction.parts) && prompt.system_instruction.parts.length) {
body.systemInstruction = prompt.system_instruction;
}
@ -868,7 +883,7 @@ async function sendXaiRequest(request, response) {
bodyParams['stop'] = request.body.stop;
}
if (['grok-3-mini-beta', 'grok-3-mini-fast-beta'].includes(request.body.model)) {
if (request.body.reasoning_effort && ['grok-3-mini-beta', 'grok-3-mini-fast-beta'].includes(request.body.model)) {
bodyParams['reasoning_effort'] = request.body.reasoning_effort === 'high' ? 'high' : 'low';
}
@ -1210,6 +1225,10 @@ router.post('/generate', function (request, response) {
bodyParams['route'] = 'fallback';
}
if (request.body.reasoning_effort) {
bodyParams['reasoning'] = { effort: request.body.reasoning_effort };
}
let cachingAtDepth = getConfigValue('claude.cachingAtDepth', -1, 'number');
if (Number.isInteger(cachingAtDepth) && cachingAtDepth >= 0 && request.body.model?.startsWith('anthropic/claude-3')) {
cachingAtDepthForOpenRouterClaude(request.body.messages, cachingAtDepth);
@ -1258,7 +1277,7 @@ router.post('/generate', function (request, response) {
}
// A few of OpenAIs reasoning models support reasoning effort
if ([CHAT_COMPLETION_SOURCES.CUSTOM, CHAT_COMPLETION_SOURCES.OPENAI].includes(request.body.chat_completion_source)) {
if (request.body.reasoning_effort && [CHAT_COMPLETION_SOURCES.CUSTOM, CHAT_COMPLETION_SOURCES.OPENAI].includes(request.body.chat_completion_source)) {
if (['o1', 'o3-mini', 'o3-mini-2025-01-31', 'o4-mini', 'o4-mini-2025-04-16', 'o3', 'o3-2025-04-16'].includes(request.body.model)) {
bodyParams['reasoning_effort'] = request.body.reasoning_effort;
}

68
src/prompt-converters.js
View File

@ -3,6 +3,15 @@ import { getConfigValue, tryParse } from './util.js';
const PROMPT_PLACEHOLDER = getConfigValue('promptPlaceholder', 'Let\'s get started.');
const REASONING_EFFORT = {
auto: 'auto',
low: 'low',
medium: 'medium',
high: 'high',
min: 'min',
max: 'max',
};
/**
* @typedef {object} PromptNames
* @property {string} charName Character name
@ -356,7 +365,7 @@ export function convertCohereMessages(messages, names) {
* @param {string} model Model name
* @param {boolean} useSysPrompt Use system prompt
* @param {PromptNames} names Prompt names
* @returns {{contents: *[], system_instruction: {parts: {text: string}}}} Prompt for Google MakerSuite models
* @returns {{contents: *[], system_instruction: {parts: {text: string}[]}}} Prompt for Google MakerSuite models
*/
export function convertGooglePrompt(messages, model, useSysPrompt, names) {
const visionSupportedModelPrefix = [
@ -369,8 +378,8 @@ export function convertGooglePrompt(messages, model, useSysPrompt, names) {
];
const isMultimodal = visionSupportedModelPrefix.some(prefix => model.startsWith(prefix));
const sysPrompt = [];
let sys_prompt = '';
if (useSysPrompt) {
while (messages.length > 1 && messages[0].role === 'system') {
// Append example names if not already done by the frontend (e.g. for group chats).
@ -384,12 +393,12 @@ export function convertGooglePrompt(messages, model, useSysPrompt, names) {
messages[0].content = `${names.charName}: ${messages[0].content}`;
}
}
sys_prompt += `${messages[0].content}\n\n`;
sysPrompt.push(messages[0].content);
messages.shift();
}
}
const system_instruction = { parts: [{ text: sys_prompt.trim() }] };
const system_instruction = { parts: sysPrompt.map(text => ({ text })) };
const toolNameMap = {};
const contents = [];
@ -919,25 +928,32 @@ export function cachingAtDepthForOpenRouterClaude(messages, cachingAtDepth) {
}
/**
* Calculate the budget tokens for a given reasoning effort.
* Calculate the Claude budget tokens for a given reasoning effort.
* @param {number} maxTokens Maximum tokens
* @param {string} reasoningEffort Reasoning effort
* @param {boolean} stream If streaming is enabled
* @returns {number} Budget tokens
*/
export function calculateBudgetTokens(maxTokens, reasoningEffort, stream) {
export function calculateClaudeBudgetTokens(maxTokens, reasoningEffort, stream) {
let budgetTokens = 0;
switch (reasoningEffort) {
case 'low':
case REASONING_EFFORT.min:
budgetTokens = 1024;
break;
case REASONING_EFFORT.low:
budgetTokens = Math.floor(maxTokens * 0.1);
break;
case 'medium':
case REASONING_EFFORT.auto:
case REASONING_EFFORT.medium:
budgetTokens = Math.floor(maxTokens * 0.25);
break;
case 'high':
case REASONING_EFFORT.high:
budgetTokens = Math.floor(maxTokens * 0.5);
break;
case REASONING_EFFORT.max:
budgetTokens = Math.floor(maxTokens * 0.95);
break;
}
budgetTokens = Math.max(budgetTokens, 1024);
@ -948,3 +964,37 @@ export function calculateBudgetTokens(maxTokens, reasoningEffort, stream) {
return budgetTokens;
}
/**
* Calculate the Google budget tokens for a given reasoning effort.
* @param {number} maxTokens Maximum tokens
* @param {string} reasoningEffort Reasoning effort
* @returns {number?} Budget tokens
*/
export function calculateGoogleBudgetTokens(maxTokens, reasoningEffort) {
let budgetTokens = 0;
switch (reasoningEffort) {
case REASONING_EFFORT.auto:
return null;
case REASONING_EFFORT.min:
budgetTokens = 0;
break;
case REASONING_EFFORT.low:
budgetTokens = Math.floor(maxTokens * 0.1);
break;
case REASONING_EFFORT.medium:
budgetTokens = Math.floor(maxTokens * 0.25);
break;
case REASONING_EFFORT.high:
budgetTokens = Math.floor(maxTokens * 0.5);
break;
case REASONING_EFFORT.max:
budgetTokens = maxTokens;
break;
}
budgetTokens = Math.min(budgetTokens, 24576);
return budgetTokens;
}