Merge branch 'staging' into vision-cleanup

This commit is contained in:
equal-l2
2025-04-25 01:49:40 +09:00
5 changed files with 142 additions and 27 deletions

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;
}