mirror of
https://github.com/SilenceLurker/ST-Amily2-Chat-Optimisation.git
synced 2026-06-06 22:55:51 +00:00
324 lines
13 KiB
JavaScript
324 lines
13 KiB
JavaScript
import { getRequestHeaders } from "/script.js";
|
||
import { getContext, extension_settings } from "/scripts/extensions.js";
|
||
import { amilyHelper } from '../../../core/tavern-helper/main.js';
|
||
import Options from './Options.js';
|
||
import RequestBody from './RequestBody.js';
|
||
|
||
/**
|
||
* ModelCaller Service
|
||
* 负责执行 API 调用逻辑，旨在替换 NccsApi 及其他散乱的请求逻辑。
|
||
* 支持：标准直连、ST预设调用、伪流式聚合(防超时)、数据归一化。
|
||
*/
|
||
export default class ModelCaller {
|
||
/**
|
||
* 构造函数注入 Logger 依赖
|
||
* @param {Object} loggerDelegate - { log: (level, msg, origin, plugin) => void }
|
||
*/
|
||
constructor(loggerDelegate) {
|
||
/** @type {Object} */
|
||
this.logger = loggerDelegate;
|
||
this.defaultHeaders = {
|
||
'Content-Type': 'application/json'
|
||
};
|
||
}
|
||
|
||
/**
|
||
* 统一调用入口
|
||
* @param {string} callerName - 调用者名称（日志用）
|
||
* @param {Array} messages - 聊天消息历史
|
||
* @param {Options} options - 配置对象实例
|
||
* @returns {Promise<string>} - 返回归一化后的文本内容
|
||
*/
|
||
async call(callerName, messages, options) {
|
||
// 1. 强制类型检查
|
||
if (!(options instanceof Options)) {
|
||
const errorMsg = `[ModelCaller] Options must be instance of Options class.`;
|
||
throw new TypeError(errorMsg);
|
||
}
|
||
|
||
// 2. 逻辑中直接使用 options 属性
|
||
// 记录一下当前的流模式，方便调试
|
||
this._log('info', `API Request [${options.mode}] StreamMode: ${options.fakeStream}`, callerName);
|
||
|
||
try {
|
||
// 统一构建请求体 DTO
|
||
const requestBody = new RequestBody(messages, options);
|
||
let result;
|
||
|
||
if (options.mode === 'preset') {
|
||
result = await this._callPreset(callerName, requestBody, options);
|
||
} else {
|
||
result = await this._callDirect(callerName, requestBody, options);
|
||
}
|
||
|
||
// 如果是流式返回，result 已经是拼接好的字符串，不需要 normalize 的部分逻辑
|
||
// 但为了统一，我们还是传进去检查一下
|
||
return this._normalize(result, options.fakeStream);
|
||
} catch (error) {
|
||
this._log('error', `Request Failed: ${error.message}`, callerName);
|
||
throw error;
|
||
}
|
||
}
|
||
|
||
// 内部日志封装
|
||
_log(level, msg, plugin) {
|
||
if (this.logger?.log) {
|
||
this.logger.log(level, msg, 'ModelCaller', plugin);
|
||
}
|
||
}
|
||
|
||
// ========================================================================
|
||
// 模式一：Direct (标准直连)
|
||
// 对应 NccsApi 中的 callNccsOpenAITest
|
||
// ========================================================================
|
||
async _callDirect(callerName, requestBody, options) {
|
||
// 构建标准 OpenAI 兼容 Body
|
||
// 目标通常是 ST 的后端代理接口
|
||
const url = '/api/backends/chat-completions/generate';
|
||
const payload = requestBody.toPayload(); // 使用 DTO 生成数据
|
||
|
||
const fetchOpts = {
|
||
method: 'POST',
|
||
headers: { ...getRequestHeaders(), ...this.defaultHeaders },
|
||
body: JSON.stringify(payload)
|
||
};
|
||
|
||
return options.fakeStream
|
||
? this._fetchFakeStream(url, fetchOpts)
|
||
: this._fetchStandard(url, fetchOpts);
|
||
}
|
||
|
||
// ========================================================================
|
||
// 模式二：Preset (ST预设调用)
|
||
// 对应 NccsApi 中的 callNccsSillyTavernPreset
|
||
// ========================================================================
|
||
async _callPreset(callerName, requestBody, options) {
|
||
const context = getContext();
|
||
|
||
// 1. 记录并切换 Profile
|
||
const originalProfile = await amilyHelper.triggerSlash('/profile');
|
||
const targetProfile = context.extensionSettings?.connectionManager?.profiles?.find(p => p.id === options.presetId);
|
||
|
||
if (!targetProfile) throw new Error(`Preset ID ${options.presetId} not found`);
|
||
|
||
if (originalProfile !== targetProfile.name) {
|
||
this._log('info', `Switching profile: ${originalProfile} -> ${targetProfile.name}`, callerName);
|
||
const escapedName = targetProfile.name.replace(/"/g, '\\"');
|
||
await amilyHelper.triggerSlash(`/profile await=true "${escapedName}"`);
|
||
}
|
||
|
||
try {
|
||
// 2. 根据流式需求分流处理
|
||
if (options.fakeStream) {
|
||
// 【流式预设请求】
|
||
// 难点：ST 的 ConnectionManagerRequestService 不暴露流。
|
||
// 策略：切换 Profile 后，手动向生成接口发送请求。
|
||
const url = '/api/backends/chat-completions/generate';
|
||
|
||
// [修复]: 手动合并 Profile 中的关键参数，否则后端不会自动应用预设配置
|
||
// 提取逻辑已封装至 _buildProfilePayload
|
||
const profilePayload = this._buildProfilePayload(targetProfile);
|
||
|
||
// 合并顺序：基础Payload(msg) < Profile预设 < 显式Params覆盖
|
||
// toMinimalPayload 包含: messages, stream, max_tokens, ...params
|
||
const minimal = requestBody.toMinimalPayload();
|
||
|
||
const finalPayload = {
|
||
...profilePayload,
|
||
...minimal,
|
||
...options.params
|
||
};
|
||
|
||
const fetchOpts = {
|
||
method: 'POST',
|
||
headers: { ...getRequestHeaders(), ...this.defaultHeaders },
|
||
body: JSON.stringify(finalPayload)
|
||
};
|
||
return await this._fetchFakeStream(url, fetchOpts);
|
||
} else {
|
||
// 【非流式预设请求】
|
||
// 直接使用 ST 原生服务，最稳妥
|
||
if (!context.ConnectionManagerRequestService) throw new Error('ST Request Service unavailable');
|
||
return await context.ConnectionManagerRequestService.sendRequest(
|
||
targetProfile.id,
|
||
requestBody.messages,
|
||
options.maxTokens
|
||
);
|
||
}
|
||
|
||
} finally {
|
||
// 3. 恢复 Profile
|
||
if (originalProfile) {
|
||
try {
|
||
const current = await amilyHelper.triggerSlash('/profile');
|
||
if (originalProfile !== current) {
|
||
const escapedOriginal = originalProfile.replace(/"/g, '\\"');
|
||
await amilyHelper.triggerSlash(`/profile await=true "${escapedOriginal}"`);
|
||
}
|
||
} catch (e) {
|
||
this._log('warn', `Failed to restore profile: ${e.message}`, callerName);
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// ========================================================================
|
||
// 网络层核心
|
||
// ========================================================================
|
||
|
||
async _fetchStandard(url, opts) {
|
||
const res = await fetch(url, opts);
|
||
if (!res.ok) throw new Error(`HTTP ${res.status}`);
|
||
return await res.json();
|
||
}
|
||
|
||
// 【核心升级】：支持 SSE 解析的伪流式聚合，防 CloudFlare 超时
|
||
async _fetchFakeStream(url, opts) {
|
||
const res = await fetch(url, opts);
|
||
if (!res.ok) throw new Error(`Stream HTTP ${res.status}`);
|
||
const reader = res.body.getReader();
|
||
const decoder = new TextDecoder();
|
||
let fullContent = ""; // 用于存储最终拼接的纯文本
|
||
let buffer = ""; // 用于存储未处理完的数据片段
|
||
try {
|
||
while (true) {
|
||
const { done, value } = await reader.read();
|
||
if (done) break;
|
||
// 1. 解码当前数据包
|
||
const chunk = decoder.decode(value, { stream: true });
|
||
buffer += chunk;
|
||
|
||
// 2. 处理 SSE 格式 (data: {...})
|
||
// 以双换行符分割每一条 SSE 消息
|
||
const lines = buffer.split('\n');
|
||
|
||
// 保留最后一个可能不完整的片段在 buffer 中
|
||
buffer = lines.pop();
|
||
|
||
for (const line of lines) {
|
||
const trimmed = line.trim();
|
||
if (!trimmed || trimmed === 'data: [DONE]') continue;
|
||
|
||
if (trimmed.startsWith('data: ')) {
|
||
try {
|
||
const jsonStr = trimmed.substring(6); // 去掉 'data: '
|
||
const json = JSON.parse(jsonStr);
|
||
|
||
// 提取 delta content
|
||
const delta = json.choices?.[0]?.delta?.content;
|
||
if (delta) {
|
||
fullContent += delta;
|
||
}
|
||
} catch (e) {
|
||
// 忽略解析错误的行，防止因为个别丢包导致整个请求失败
|
||
console.warn('[ModelCaller] SSE Parse Error:', e);
|
||
}
|
||
}
|
||
}
|
||
}
|
||
} finally {
|
||
reader.releaseLock();
|
||
}
|
||
|
||
// 如果 fullContent 是空的，说明可能服务端根本没返回 SSE 格式，而是直接返回了纯文本或 JSON
|
||
// 这种情况下尝试降级处理
|
||
if (!fullContent && buffer) {
|
||
try {
|
||
const json = JSON.parse(buffer);
|
||
return json; // 是标准 JSON
|
||
} catch {
|
||
return buffer; // 是纯文本
|
||
}
|
||
}
|
||
|
||
return fullContent;
|
||
}
|
||
|
||
// ========================================================================
|
||
// 数据归一化
|
||
// ========================================================================
|
||
|
||
_normalize(data, isFromStream = false) {
|
||
// 如果是从流式聚合来的，它已经是一个纯字符串了，直接返回
|
||
if (isFromStream && typeof data === 'string') {
|
||
return data;
|
||
}
|
||
|
||
// 如果是 JSON 字符串则解析
|
||
if (typeof data === 'string') {
|
||
try { data = JSON.parse(data); } catch (e) { return data; }
|
||
}
|
||
|
||
// 处理 OpenAI 格式
|
||
if (data?.choices?.[0]?.message?.content) {
|
||
return data.choices[0].message.content.trim();
|
||
}
|
||
|
||
// 处理常规 content 格式
|
||
if (data?.content) {
|
||
return data.content.trim();
|
||
}
|
||
|
||
// Fallback
|
||
return typeof data === 'object' ? JSON.stringify(data) : data;
|
||
}
|
||
|
||
/**
|
||
* 辅助方法：从 Profile 对象中提取标准生成参数
|
||
* 严格复刻 SillyTavern 原始 Payload 逻辑
|
||
*/
|
||
_buildProfilePayload(targetProfile) {
|
||
const context = getContext();
|
||
|
||
// 1. 基础克隆
|
||
const payload = { ...targetProfile };
|
||
|
||
// 2. 注入运行时元数据 (这是旧版能通的关键，包含用户/角色名等)
|
||
payload.user_name = context.name1 || 'User';
|
||
payload.char_name = context.name2 || 'AI';
|
||
payload.group_names = []; // 暂不处理群组
|
||
payload.use_sysprompt = true;
|
||
payload.type = 'quiet';
|
||
payload.custom_prompt_post_processing = payload.custom_prompt_post_processing || 'strict';
|
||
|
||
// 3. 规范化模型字段
|
||
if (!payload.model) {
|
||
payload.model = payload.openai_model || payload.claude_model || payload.mistral_model || '';
|
||
}
|
||
|
||
// 4. 精准对齐 URL 映射 (解决 403/400 错误的核心)
|
||
const rawUrl = payload['api-url'] || payload['api_url'] || payload.custom_url || payload.url;
|
||
if (rawUrl) {
|
||
// 如果 Source 是 custom，严格遵循旧版：custom_url 有值，reverse_proxy 为空
|
||
if (payload.chat_completion_source === 'custom') {
|
||
payload.custom_url = rawUrl;
|
||
payload.reverse_proxy = payload.reverse_proxy || '';
|
||
} else {
|
||
// 如果是 openai，则填充 reverse_proxy
|
||
payload.reverse_proxy = rawUrl;
|
||
payload.custom_url = rawUrl;
|
||
}
|
||
// 兼容性修补
|
||
payload.zai_endpoint = rawUrl;
|
||
payload.vertexai_region = rawUrl;
|
||
}
|
||
|
||
// 5. 补全采样参数 (严格对齐 UI 当前状态)
|
||
const globalGenSettings = extension_settings.text_generation || {};
|
||
const fields = ['temperature', 'max_tokens', 'top_p', 'top_k', 'min_p', 'frequency_penalty', 'presence_penalty', 'repetition_penalty'];
|
||
fields.forEach(field => {
|
||
if (payload[field] === undefined) {
|
||
payload[field] = globalGenSettings[field] ?? (field === 'temperature' ? 1 : 0);
|
||
}
|
||
});
|
||
|
||
// 6. 确保 Source 存在且不被错误覆盖
|
||
if (!payload.chat_completion_source) {
|
||
payload.chat_completion_source = 'openai';
|
||
}
|
||
|
||
return payload;
|
||
}
|
||
}
|
||
|