mirror of
https://github.com/Wx-2025/ST-Amily2-Chat-Optimisation.git
synced 2026-06-13 11:35:51 +00:00
release: v2.2.6 [2026-06-13 01:02:05]
### 新功能 - **翰林院向量化质量升级**： - **边界感知切块**：替换四个来源（聊天记录/小说/世界书/手动）的纯字符硬切——优先在段落边界断开，其次句末标点（含中文引号闭合），极端长串才硬切；句子/对话不再被拦腰截断，embedding 质量同步受益。仅影响新录入，已有向量无需重建 - **注入时序重排**：检索结果注入提示词前按时序重排（聊天记录按楼层、小说按卷/章/节——中文数字章节号可解析），rerank 只决定"选哪些块"，不再决定呈现顺序；修复"不打不相识的剧情之后紧跟关系亲密"这类因按相关度排序导致的认知时间错乱 - **断层提示**：聊天记录相邻块楼层跳跃时自动插入"与上文相隔约 N 楼，并非连续发生"提示行，消除中间剧情缺失造成的割裂感 - **时间标识**：新录入的聊天记录块在来源标识中带上消息发送时间（ST 向量存储不持久化元数据，时间必须写入块文本才能在检索后取回；旧格式块兼容解析） - **记忆块工作流（memory-blocks）**：剧情优化新增"自定义记忆块"体系——占位符驱动的并发工作流框架 - 在剧情优化面板「匹配替换 (sulv)」下方可增删自定义块：每个块定义一个占位符，执行剧情优化时主/拦截提示词中的占位符会被块的产出替换 - **静态块**：直接输出固定内容；**AI 调用块**：用所选 API 功能槽独立请求一次，把回复（或其中指定 `<标签>` 的内容）作为替换值 - 原有 sulv1-4 速率占位符迁入同一框架，行为与旧版逐字节一致 - 块定义为纯 JSON、随设置持久化，为后续导入导出与战斗系统接入预留扩展点 - 框架层新增**顺序拼接式 Chain**（`composeChain`）：与占位符替换并列的第二种组合范式——同链的块并发执行后按 `order` 排序、以 `separator` 拼接并可选 `header/footer` 包裹，产出一个完整注入块；为记忆注入合成块与战斗系统"底部战报块"预留的承载结构，本版本暂无 UI 入口 - **API 连接配置**： - 角色世界书（cwb）与一键生卡（autoCharCard）纳入旧配置自动迁移：老用户首次加载会把旧 URL / Key / 模型自动迁移为连接配置并分配槽位（一键生卡仅在规划者与执行者配置一致或规划者为空时迁移，避免悄悄改变行为） - **profile 已分配时参数控件 informational 化**：主面板 / 并发剧情优化 / 角色世界书 / 术语表的温度、maxTokens 控件在槽位分配 profile 后自动禁用并显示"由连接配置控制"提示，消除"改了没效果"的用户陷阱 - **profile 状态卡新增"本设备无 Key"警示**：API Key 仅保存在最初填写它的设备/浏览器上（安全设计，不随云端设置同步），换设备后状态卡会直接亮出警示徽标，不必等到调用报错才发现 ### 修复 - **独立聊天记忆从摆设变真功能（原作遗留坑）**：此前向量数据"随卡不随聊天"——开启"独立聊天记忆"后录入仍存进角色库、查询却去查一个从未被写入过的聊天集合、计数恒为 0，整体静默失效。现已重构为聊天级分桶： - 独立模式下，聊天记录类向量按当前聊天隔离存储与检索，同一张卡开多个聊天（不同剧情线）的记忆互不污染 - 小说 / 世界书 / 手动录入属于"知识"，仍随角色卡跨聊天共享；全局库不受影响 - 知识管理列表为聊天专属库显示"聊天级"徽标；聊天级库禁止移动到全局 - 统一模式（默认关闭独立记忆）的存量数据与行为完全不变 - 已知限制：聊天专属记忆跟随聊天文件，重命名聊天文件会使其失联（与 ST 官方向量扩展同等限制） - **超级排序截断顺序修正**：开启"超级排序"时，时序重排发生在 top_n 截断之前，导致保留的是"时序最早"而非"最相关"的块，检索结果长期偏向最旧的聊天记录。现改为先按相关度截取 top_n、再做时序排序 - **翰林院向量化失败（"向量化块数量不识别"反馈）**： - 一次性清洗 profile-sync 历史污染：`retrieval/rerank.apiKey` 中的掩码占位符在持久层根治（此前仅读取侧防御）；`apiEndpoint` / `rerank.apiMode` 的非法值（如被旧版写入的空字符串）归一化为 `custom` - 修复 `apiEndpoint` 为空/非法时请求被硬定向到 `api.openai.com`、无视用户自定义 URL 的问题（CSP 拦截 / 401 的元凶） - 修复**本地代理（LM Studio/Ollama）模式**自始就缺少 URL 分支、同样被错误定向到 openai.com 的问题 - API 模式下拉补全 `OpenAI 官方` / `Azure` 选项；默认 API 模式改为 `custom`（与默认 URL 配套），新用户不再因选项缺失导致首次保存写入空值 - profile-sync 给下拉框赋不存在选项值的污染源头修复（影响所有模块面板，不止翰林院） - **Rerank "API Key 未提供"报错升级**：当原因是"连接配置在本设备没有可用 Key"时，报错会直接说明 Key 的设备本地性并指引到 API 连接配置重新填写（向量化 Google 直连、获取模型列表同步处理） - **旧配置迁移**：一键生卡迁移时排除掩码占位符，避免把历史污染的假 Key 迁入新连接配置 - **超级记忆稳定性专项**（针对"工作不大稳定"反馈，4 处根因一次修复）： - **切聊天竞态污染**：CHAT_CHANGED 时超级记忆立即全量同步，而表格系统延迟 100ms 才加载新聊天的表格，导致【旧聊天】的表格内容被写进【新角色】的记忆世界书；两边表名不同时旧表条目无 GC 兜底会**永久残留**（"记忆串台"元凶）。现 CHAT_CHANGED 只确保世界书存在，新状态同步交由 `loadTables()` 完成后的自动推送，单次且时序正确 - **死代码双轨存储拆除**：`saveStateToMetadata` / `tryRestoreStateFromMetadata` 把表格状态写到 `msg.metadata`——该字段非 ST 持久化位（同 v2.2.5 二次填表修过的坑），写入即蒸发、恢复永远为空，且每次同步还白调一次 `saveChat()`。整条链路删除，表格状态唯一信源为表格系统的 `msg.extra.amily2_tables_data` - **`awaitSync()` 穿透**：同步队列正忙时 `pushUpdate` 会用一个立即 resolve 的空 Promise 覆盖 `_syncPromise`，Pipeline Stage 4 等待形同虚设、后续阶段在同步未完成时被放行。现忙时不覆盖，正在运行的 drain 循环自然吃掉新入队项 - **开关打开不生效**：启动时若总开关为关，初始化早退且不注册监听器；此后在 UI 勾选开关只写设置，超级记忆直到刷新页面前都是死的。现勾选即触发初始化（幂等） - 附带：`forceSyncAll` 的表格角色推断改为复用 `events-schema.inferTableRole`，消除两处重复逻辑漂移风险；每次切聊天的双倍全量同步（restore 路径一次 + 显式一次）随死代码移除归一 ### 重构 - 表格核心 `manager.js` 瘦身（约 1050 → 600 行）：19 个 UI 突变操作拆分至 `actions/ui-mutations.js`，SuperMemory 事件分发拆分至 `events-dispatch.js`；全部经 re-export 保持兼容，外部调用路径零改动 - 角色世界书最后 2 处散乱的厂商 URL 判断迁移至 `detectVendor` 统一入口，业务路径上不再有硬编码的 URL substring 判断
This commit is contained in:
Jenkins CI
@@ -152,6 +152,7 @@ function initialize() {
|
||||
return;
|
||||
}
|
||||
migrateLegacyRagSettings();
|
||||
sanitizeProfilePollution();
|
||||
settings = getSettings();
|
||||
if (!window.hanlinyuanRagProcessor) {
|
||||
window.hanlinyuanRagProcessor = {};
|
||||
@@ -219,20 +220,27 @@ async function ingestTextToHanlinyuan(text, source = 'manual', metadata = {}, pr
|
||||
break;
|
||||
}
|
||||
|
||||
// 独立聊天记忆模式：聊天记录类向量按聊天分桶（剧情线隔离），
|
||||
// 其余来源（小说/世界书/手动）属于"知识"，仍随角色卡共享
|
||||
const independentChatId = (source === 'chat_history' && settings.retrieval.independentChatMemoryEnabled)
|
||||
? getChatId()
|
||||
: null;
|
||||
|
||||
const existingKbs = Object.values(getKnowledgeBases());
|
||||
const foundKb = existingKbs.find(kb => kb.name === kbName);
|
||||
// 同名合并需限定在同一聊天命名空间内，避免独立模式下不同聊天的同名楼层段互相串库
|
||||
const foundKb = existingKbs.find(kb => kb.name === kbName && (kb.chatId ?? null) === independentChatId);
|
||||
|
||||
if (foundKb) {
|
||||
taskId = foundKb.id;
|
||||
logCallback(`[翰林院-核心] 检测到同名知识库 "${kbName}"，将数据合并入库。`, 'info');
|
||||
} else {
|
||||
logCallback(`[翰林院-核心] 准备为任务 "${kbName}" 创建专属知识库...`, 'info');
|
||||
const newKb = addKnowledgeBase(kbName, source);
|
||||
const newKb = addKnowledgeBase(kbName, source, independentChatId);
|
||||
taskId = newKb.id;
|
||||
}
|
||||
|
||||
|
||||
const charId = getCharacterStableId();
|
||||
const collectionId = `${charId}_${taskId}`;
|
||||
const collectionId = independentChatId ? `${independentChatId}_${taskId}` : `${charId}_${taskId}`;
|
||||
logCallback(`[翰林院-核心] 已创建并锁定知识库: ${kbName} (集合ID: ${collectionId})`, 'success');
|
||||
logCallback(`[翰林院-核心] 已锁定忆识宝库ID: ${collectionId}`, 'info');
|
||||
|
||||
@@ -410,6 +418,49 @@ function migrateLegacyRagSettings() {
|
||||
saveSettingsDebounced();
|
||||
}
|
||||
|
||||
/**
|
||||
* 一次性清洗 profile-sync 历史污染（2.2.5 之前的版本遗留）。
|
||||
*
|
||||
* 旧版 saveSettingsFromUI 会把被 Profile 接管的隐藏字段值写回 settings：
|
||||
* - apiKey 被写成掩码 '••••••••'（rag-api 已有读侧防御，这里根治持久层）
|
||||
* - apiEndpoint 的 select 被 _fillLegacyFields 赋了不存在的 option 值
|
||||
* （profile.provider 如 'custom_oai'）后 value 变 ''，'' 被写回 settings；
|
||||
* '' 在 getApiEndpointUrl 落 default 分支，请求被错误定向 → 向量化全失败
|
||||
*
|
||||
* 2.2.5 修复了"继续污染"，本函数清理已污染的存量数据。
|
||||
*/
|
||||
function sanitizeProfilePollution() {
|
||||
const s = getSettings();
|
||||
const MASKED = '••••••••';
|
||||
let cleaned = [];
|
||||
|
||||
if (s.retrieval?.apiKey === MASKED) {
|
||||
s.retrieval.apiKey = '';
|
||||
cleaned.push('retrieval.apiKey 掩码');
|
||||
}
|
||||
if (s.rerank?.apiKey === MASKED) {
|
||||
s.rerank.apiKey = '';
|
||||
cleaned.push('rerank.apiKey 掩码');
|
||||
}
|
||||
|
||||
// 合法值与 UI select 选项及 rag-api 的 switch 分支保持一致
|
||||
const validEndpoints = ['custom', 'google_direct', 'local_proxy', 'openai', 'azure'];
|
||||
if (s.retrieval && !validEndpoints.includes(s.retrieval.apiEndpoint)) {
|
||||
cleaned.push(`retrieval.apiEndpoint 非法值 "${s.retrieval.apiEndpoint}"`);
|
||||
s.retrieval.apiEndpoint = 'custom';
|
||||
}
|
||||
const validRerankModes = ['custom', 'local_proxy'];
|
||||
if (s.rerank && !validRerankModes.includes(s.rerank.apiMode)) {
|
||||
cleaned.push(`rerank.apiMode 非法值 "${s.rerank.apiMode}"`);
|
||||
s.rerank.apiMode = 'custom';
|
||||
}
|
||||
|
||||
if (cleaned.length > 0) {
|
||||
console.warn(`[翰林院] 已清洗 profile-sync 历史污染字段: ${cleaned.join('、')}`);
|
||||
saveSettings();
|
||||
}
|
||||
}
|
||||
|
||||
function showNotification(message, type = 'info') {
|
||||
toastr[type](message);
|
||||
}
|
||||
@@ -431,6 +482,71 @@ function getTagForSource(source) {
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* 边界感知切分：把 content 切成不超过 chunkSize 的片段，尽量在自然边界断开。
|
||||
*
|
||||
* 三级回退策略（替代旧的纯字符硬切，避免句子/对话被拦腰截断）：
|
||||
* 1. 段落边界（最后一个换行符）
|
||||
* 2. 句末边界（。！？!?… 及其后跟随的闭合引号/括号）
|
||||
* 3. 都找不到（极端长串）才硬切
|
||||
* 边界切点过于靠前（< 40% 块长）时视为无效，降级到下一策略——防止
|
||||
* 一个超长段落开头的短句导致块碎片化。
|
||||
*
|
||||
* @param {string} content
|
||||
* @param {number} chunkSize - 单块最大字符数
|
||||
* @param {number} overlap - 相邻块重叠字符数（语义衔接），从上一块尾部回看
|
||||
* @returns {string[]}
|
||||
*/
|
||||
function splitBySemanticBoundary(content, chunkSize, overlap) {
|
||||
const pieces = [];
|
||||
if (!content || chunkSize <= 0) return pieces;
|
||||
|
||||
const minCut = Math.floor(chunkSize * 0.4);
|
||||
const sentenceEndRegex = /[。！？!?…][”"』」)）】]?/g;
|
||||
|
||||
let pos = 0;
|
||||
while (pos < content.length) {
|
||||
let end = Math.min(pos + chunkSize, content.length);
|
||||
|
||||
if (end < content.length) {
|
||||
const slice = content.substring(pos, end);
|
||||
|
||||
// 1. 段落边界：最后一个换行（切点含换行符本身）
|
||||
let cut = slice.lastIndexOf('\n') + 1;
|
||||
|
||||
// 2. 段落边界无效时找最后一个句末边界
|
||||
if (cut <= minCut) {
|
||||
let lastSentenceEnd = -1;
|
||||
sentenceEndRegex.lastIndex = 0;
|
||||
let m;
|
||||
while ((m = sentenceEndRegex.exec(slice)) !== null) {
|
||||
lastSentenceEnd = m.index + m[0].length;
|
||||
}
|
||||
if (lastSentenceEnd > minCut) cut = lastSentenceEnd;
|
||||
}
|
||||
|
||||
// 3. 有效边界则收缩切点，否则保持硬切
|
||||
if (cut > minCut) end = pos + cut;
|
||||
}
|
||||
|
||||
const piece = content.substring(pos, end);
|
||||
if (piece.trim().length > 0) pieces.push(piece);
|
||||
|
||||
if (end >= content.length) break;
|
||||
// overlap 回看；Math.max 防止 overlap >= 块长时死循环
|
||||
pos = Math.max(end - overlap, pos + 1);
|
||||
}
|
||||
return pieces;
|
||||
}
|
||||
|
||||
/** 把 ISO/任意时间值格式化为写入块 prefix 的紧凑标识（不含逗号，便于正则反解） */
|
||||
function formatChunkTimeLabel(timestamp) {
|
||||
const d = new Date(timestamp);
|
||||
if (isNaN(d.getTime())) return '';
|
||||
const pad = n => String(n).padStart(2, '0');
|
||||
return `${d.getFullYear()}-${pad(d.getMonth() + 1)}-${pad(d.getDate())} ${pad(d.getHours())}:${pad(d.getMinutes())}`;
|
||||
}
|
||||
|
||||
function splitIntoChunks(text, source, metadata = {}) {
|
||||
switch (source) {
|
||||
case 'novel':
|
||||
@@ -465,30 +581,22 @@ function _chunkForNovel(text, metadata) {
|
||||
function processBuffer() {
|
||||
if (contentBuffer.length === 0) return;
|
||||
const content = contentBuffer.join('\n');
|
||||
let start = 0;
|
||||
let section = 1;
|
||||
while (start < content.length) {
|
||||
const end = Math.min(start + chunkSize, content.length);
|
||||
const chunkText = content.substring(start, end);
|
||||
if (chunkText.trim().length > 0) {
|
||||
const chunkMetadata = {
|
||||
source: 'novel',
|
||||
sourceName: sourceName,
|
||||
timestamp: new Date().toISOString(),
|
||||
globalIndex: globalChunkIndex++,
|
||||
volume: currentVolumeTitle,
|
||||
chapter: currentChapterTitle,
|
||||
section: section,
|
||||
};
|
||||
const tagName = getTagForSource('novel');
|
||||
const prefix = `[来源: ${sourceName}, ${currentVolumeTitle}, ${currentChapterTitle}, 第${section}节]`;
|
||||
const wrappedText = `<${tagName}>\n${prefix}\n${chunkText}\n</${tagName}>`;
|
||||
allChunks.push({ text: wrappedText, metadata: chunkMetadata });
|
||||
section++;
|
||||
}
|
||||
start += (chunkSize - overlap);
|
||||
if (start >= content.length) break;
|
||||
}
|
||||
const tagName = getTagForSource('novel');
|
||||
splitBySemanticBoundary(content, chunkSize, overlap).forEach((chunkText, idx) => {
|
||||
const section = idx + 1;
|
||||
const chunkMetadata = {
|
||||
source: 'novel',
|
||||
sourceName: sourceName,
|
||||
timestamp: new Date().toISOString(),
|
||||
globalIndex: globalChunkIndex++,
|
||||
volume: currentVolumeTitle,
|
||||
chapter: currentChapterTitle,
|
||||
section: section,
|
||||
};
|
||||
const prefix = `[来源: ${sourceName}, ${currentVolumeTitle}, ${currentChapterTitle}, 第${section}节]`;
|
||||
const wrappedText = `<${tagName}>\n${prefix}\n${chunkText}\n</${tagName}>`;
|
||||
allChunks.push({ text: wrappedText, metadata: chunkMetadata });
|
||||
});
|
||||
contentBuffer = [];
|
||||
}
|
||||
|
||||
@@ -508,11 +616,9 @@ function _chunkForNovel(text, metadata) {
|
||||
processBuffer();
|
||||
|
||||
if (allChunks.length === 0 && text.length > 0) {
|
||||
let start = 0;
|
||||
let section = 1;
|
||||
while (start < text.length) {
|
||||
const end = Math.min(start + chunkSize, text.length);
|
||||
const chunkText = text.substring(start, end);
|
||||
const tagName = getTagForSource('novel');
|
||||
splitBySemanticBoundary(text, chunkSize, overlap).forEach((chunkText, idx) => {
|
||||
const section = idx + 1;
|
||||
const chunkMetadata = {
|
||||
source: 'novel',
|
||||
sourceName: sourceName,
|
||||
@@ -522,13 +628,10 @@ function _chunkForNovel(text, metadata) {
|
||||
chapter: "第1章",
|
||||
section: section,
|
||||
};
|
||||
const tagName = getTagForSource('novel');
|
||||
const prefix = `[来源: ${sourceName}, 第1卷, 第1章, 第${section}节]`;
|
||||
const wrappedText = `<${tagName}>\n${prefix}\n${chunkText}\n</${tagName}>`;
|
||||
allChunks.push({ text: wrappedText, metadata: chunkMetadata });
|
||||
section++;
|
||||
start += (chunkSize - overlap);
|
||||
}
|
||||
});
|
||||
}
|
||||
return allChunks;
|
||||
}
|
||||
@@ -540,15 +643,15 @@ function _chunkForChatHistory(text, metadata) {
|
||||
const allChunks = [];
|
||||
if (!text || chunkSize <= 0) return allChunks;
|
||||
|
||||
let part = 1;
|
||||
let start = 0;
|
||||
// 时间写进 prefix 才能在检索后被反解回来（ST 向量存储不持久化 metadata）
|
||||
const timeLabel = formatChunkTimeLabel(timestamp);
|
||||
const tagName = getTagForSource('chat_history');
|
||||
|
||||
while (start < text.length) {
|
||||
const end = Math.min(start + chunkSize, text.length);
|
||||
const chunkText = text.substring(start, end);
|
||||
|
||||
const prefix = `[来源: 聊天记录, 楼层: #${floor}, 第${part}部分]`;
|
||||
const tagName = getTagForSource('chat_history');
|
||||
splitBySemanticBoundary(text, chunkSize, overlap).forEach((chunkText, idx) => {
|
||||
const part = idx + 1;
|
||||
const prefix = timeLabel
|
||||
? `[来源: 聊天记录, 楼层: #${floor}, 时间: ${timeLabel}, 第${part}部分]`
|
||||
: `[来源: 聊天记录, 楼层: #${floor}, 第${part}部分]`;
|
||||
const wrappedText = `<${tagName}>\n${prefix}\n${chunkText}\n</${tagName}>`;
|
||||
|
||||
allChunks.push({
|
||||
@@ -562,11 +665,7 @@ function _chunkForChatHistory(text, metadata) {
|
||||
timestamp: timestamp,
|
||||
}
|
||||
});
|
||||
|
||||
part++;
|
||||
start += (chunkSize - overlap);
|
||||
if (start >= text.length) break;
|
||||
}
|
||||
});
|
||||
return allChunks;
|
||||
}
|
||||
|
||||
@@ -577,15 +676,11 @@ function _chunkForLorebook(text, metadata) {
|
||||
const allChunks = [];
|
||||
if (!text || chunkSize <= 0) return allChunks;
|
||||
|
||||
let part = 1;
|
||||
let start = 0;
|
||||
const tagName = getTagForSource('lorebook');
|
||||
|
||||
while (start < text.length) {
|
||||
const end = Math.min(start + chunkSize, text.length);
|
||||
const chunkText = text.substring(start, end);
|
||||
|
||||
splitBySemanticBoundary(text, chunkSize, overlap).forEach((chunkText, idx) => {
|
||||
const part = idx + 1;
|
||||
const prefix = `[来源: ${bookName}, 条目: ${entryName}, 第${part}部分]`;
|
||||
const tagName = getTagForSource('lorebook');
|
||||
const wrappedText = `<${tagName}>\n${prefix}\n${chunkText}\n</${tagName}>`;
|
||||
|
||||
allChunks.push({
|
||||
@@ -599,11 +694,7 @@ function _chunkForLorebook(text, metadata) {
|
||||
timestamp: new Date().toISOString(),
|
||||
}
|
||||
});
|
||||
|
||||
part++;
|
||||
start += (chunkSize - overlap);
|
||||
if (start >= text.length) break;
|
||||
}
|
||||
});
|
||||
return allChunks;
|
||||
}
|
||||
|
||||
@@ -615,16 +706,12 @@ function _chunkForManual(text, metadata) {
|
||||
if (!text || chunkSize <= 0) return allChunks;
|
||||
|
||||
const timestamp = new Date();
|
||||
const readableTime = timestamp.toLocaleString('zh-CN');
|
||||
let part = 1;
|
||||
let start = 0;
|
||||
const readableTime = formatChunkTimeLabel(timestamp);
|
||||
const tagName = getTagForSource('manual');
|
||||
|
||||
while (start < text.length) {
|
||||
const end = Math.min(start + chunkSize, text.length);
|
||||
const chunkText = text.substring(start, end);
|
||||
|
||||
splitBySemanticBoundary(text, chunkSize, overlap).forEach((chunkText, idx) => {
|
||||
const part = idx + 1;
|
||||
const prefix = `[来源: ${sourceName}, 向量化录入时间: ${readableTime}, 第${part}部分]`;
|
||||
const tagName = getTagForSource('manual');
|
||||
const wrappedText = `<${tagName}>\n${prefix}\n${chunkText}\n</${tagName}>`;
|
||||
|
||||
allChunks.push({
|
||||
@@ -636,11 +723,7 @@ function _chunkForManual(text, metadata) {
|
||||
timestamp: timestamp.toISOString(),
|
||||
}
|
||||
});
|
||||
|
||||
part++;
|
||||
start += (chunkSize - overlap);
|
||||
if (start >= text.length) break;
|
||||
}
|
||||
});
|
||||
return allChunks;
|
||||
}
|
||||
|
||||
@@ -708,7 +791,13 @@ function getKnowledgeBases() {
|
||||
return { ...globalBases, ...localBases };
|
||||
}
|
||||
|
||||
function addKnowledgeBase(name, source = 'manual') {
|
||||
/**
|
||||
* @param {string} name
|
||||
* @param {string} source
|
||||
* @param {string|null} chatId - 非空时该库为"聊天级"：向量集合按 `${chatId}_${taskId}`
|
||||
* 命名空间隔离（独立聊天记忆模式下的聊天记录库），查询时只对该聊天可见
|
||||
*/
|
||||
function addKnowledgeBase(name, source = 'manual', chatId = null) {
|
||||
if (!name || !name.trim()) {
|
||||
throw new Error('知识库名称不能为空');
|
||||
}
|
||||
@@ -721,17 +810,28 @@ function addKnowledgeBase(name, source = 'manual') {
|
||||
name: name.trim(),
|
||||
enabled: true,
|
||||
createdAt: new Date().toISOString(),
|
||||
owner: charId,
|
||||
source: source,
|
||||
owner: charId,
|
||||
source: source,
|
||||
...(chatId ? { chatId } : {}),
|
||||
};
|
||||
|
||||
bases[taskId] = newBase;
|
||||
saveSettings();
|
||||
|
||||
console.log(`[翰林院-核心] 已为角色 ${charId} 添加新知识库: ${name} (ID: ${taskId})`);
|
||||
|
||||
console.log(`[翰林院-核心] 已为角色 ${charId} 添加新知识库: ${name} (ID: ${taskId}${chatId ? `, 聊天级: ${chatId}` : ''})`);
|
||||
return newBase;
|
||||
}
|
||||
|
||||
/**
|
||||
* 计算知识库的向量集合 ID（单一事实来源）。
|
||||
* 聊天级库（kb.chatId）按聊天命名空间，其余按 owner/角色命名空间。
|
||||
*/
|
||||
function getKbCollectionId(kb, scope = 'local') {
|
||||
if (kb.chatId) return `${kb.chatId}_${kb.id}`;
|
||||
if (scope === 'global') return `${kb.owner || GLOBAL_SCOPE_ID}_${kb.id}`;
|
||||
return `${getCharacterStableId()}_${kb.id}`;
|
||||
}
|
||||
|
||||
async function removeKnowledgeBase(taskId, scope) {
|
||||
const charId = getCharacterStableId();
|
||||
const bases = scope === 'global' ? getGlobalKnowledgeBases() : getLocalKnowledgeBases();
|
||||
@@ -743,9 +843,8 @@ async function removeKnowledgeBase(taskId, scope) {
|
||||
return;
|
||||
}
|
||||
|
||||
const ownerId = scope === 'global' ? (base.owner || GLOBAL_SCOPE_ID) : charId;
|
||||
const collectionIdToPurge = `${ownerId}_${taskId}`;
|
||||
|
||||
const collectionIdToPurge = getKbCollectionId(base, scope);
|
||||
|
||||
console.log(`[翰林院-核心] 准备删除知识库 ${taskId}，将清空集合: ${collectionIdToPurge}`);
|
||||
|
||||
const purged = await purgeStorage(collectionIdToPurge);
|
||||
@@ -792,30 +891,38 @@ async function queryVectors(queryText, options = {}) {
|
||||
}
|
||||
else if (settings.retrieval.independentChatMemoryEnabled) {
|
||||
console.log('[翰林院-日志] 独立聊天记忆模式开启...');
|
||||
|
||||
|
||||
const chatId = getChatId();
|
||||
if (chatId) {
|
||||
console.log(`[翰林院-日志] 添加当前聊天宝库: ${chatId}`);
|
||||
basesToQuery.push({ id: chatId, name: `当前聊天 (${chatId})`, scope: 'chat' });
|
||||
} else {
|
||||
console.warn('[翰林院-日志] 无法获取当前聊天ID，跳过聊天宝库。');
|
||||
if (!chatId) {
|
||||
console.warn('[翰林院-日志] 无法获取当前聊天ID，聊天级知识库将被跳过。');
|
||||
}
|
||||
|
||||
const globalBases = getGlobalKnowledgeBases();
|
||||
const enabledGlobalBases = Object.values(globalBases).filter(b => b.enabled);
|
||||
// 本地库过滤规则：知识类库（无 chatId）照常可查；
|
||||
// 聊天级库（有 chatId）只对所属聊天可见——这就是"独立"的含义
|
||||
const localBases = Object.values(getLocalKnowledgeBases())
|
||||
.filter(b => b.enabled && (!b.chatId || b.chatId === chatId));
|
||||
if (localBases.length > 0) {
|
||||
const chatScoped = localBases.filter(b => b.chatId).length;
|
||||
console.log(`[翰林院-日志] 添加 ${localBases.length} 个本地知识库（其中 ${chatScoped} 个为当前聊天专属）。`);
|
||||
basesToQuery.push(...localBases.map(b => ({ ...b, scope: b.chatId ? 'chat' : 'local' })));
|
||||
}
|
||||
|
||||
const enabledGlobalBases = Object.values(getGlobalKnowledgeBases()).filter(b => b.enabled);
|
||||
if (enabledGlobalBases.length > 0) {
|
||||
console.log(`[翰林院-日志] 添加 ${enabledGlobalBases.length} 个已启用的全局知识库。`);
|
||||
basesToQuery.push(...enabledGlobalBases.map(b => ({ ...b, scope: 'global' })));
|
||||
}
|
||||
}
|
||||
}
|
||||
else {
|
||||
console.log('[翰林院-日志] 统一角色卡模式开启...');
|
||||
const localBases = getLocalKnowledgeBases();
|
||||
const globalBases = getGlobalKnowledgeBases();
|
||||
const enabledLocalBases = Object.values(localBases).filter(b => b.enabled);
|
||||
const enabledGlobalBases = Object.values(globalBases).filter(b => b.enabled);
|
||||
|
||||
basesToQuery.push(...enabledLocalBases.map(b => ({ ...b, scope: 'local' })));
|
||||
|
||||
// 聊天级库（独立模式期间产生）在统一模式下也可见，但需用 'chat' scope
|
||||
// 才能拼出正确的集合 ID（${chatId}_${taskId}）
|
||||
basesToQuery.push(...enabledLocalBases.map(b => ({ ...b, scope: b.chatId ? 'chat' : 'local' })));
|
||||
basesToQuery.push(...enabledGlobalBases.map(b => ({ ...b, scope: 'global' })));
|
||||
|
||||
if (basesToQuery.length === 0) {
|
||||
@@ -879,7 +986,9 @@ async function _executeQueryForBase(base, queryText, queryEmbedding = null) {
|
||||
collectionId = await getDynamicCollectionId();
|
||||
break;
|
||||
case 'chat':
|
||||
collectionId = base.id;
|
||||
// 聊天级库：${chatId}_${taskId} 命名空间（独立聊天记忆）。
|
||||
// 旧语义的裸 chatId 集合从未被任何录入路径写入过，无存量兼容负担
|
||||
collectionId = base.chatId ? `${base.chatId}_${base.id}` : base.id;
|
||||
break;
|
||||
case 'global':
|
||||
const ownerId = base.owner || GLOBAL_SCOPE_ID;
|
||||
@@ -945,10 +1054,12 @@ async function _executeQueryForBase(base, queryText, queryEmbedding = null) {
|
||||
switch (sourceTag) {
|
||||
case '聊天记录':
|
||||
newMetadata.source = 'chat_history';
|
||||
const chatMatch = item.text.match(/楼层:\s*#(\d+),\s*第(\d+)部分/);
|
||||
if (chatMatch && chatMatch[1] && chatMatch[2]) {
|
||||
// 时间段为可选：兼容旧格式 [楼层: #X, 第Y部分] 与新格式 [楼层: #X, 时间: ..., 第Y部分]
|
||||
const chatMatch = item.text.match(/楼层:\s*#(\d+)(?:,\s*时间:\s*([^,\]]+))?,\s*第(\d+)部分/);
|
||||
if (chatMatch && chatMatch[1] && chatMatch[3]) {
|
||||
newMetadata.floor = parseInt(chatMatch[1], 10);
|
||||
newMetadata.part = parseInt(chatMatch[2], 10);
|
||||
if (chatMatch[2]) newMetadata.timeLabel = chatMatch[2].trim();
|
||||
newMetadata.part = parseInt(chatMatch[3], 10);
|
||||
newMetadata.sourceName = `聊天记录 #${newMetadata.floor}`;
|
||||
}
|
||||
break;
|
||||
@@ -1051,43 +1162,40 @@ async function getVectorCount(taskId = null, scope = 'local') {
|
||||
console.warn(`[翰林院-计数] 在作用域 '${scope}' 中未找到ID为 ${taskId} 的知识库。`);
|
||||
return 0;
|
||||
}
|
||||
const ownerId = scope === 'global' ? (base.owner || GLOBAL_SCOPE_ID) : charId;
|
||||
const collectionId = `${ownerId}_${taskId}`;
|
||||
return await countVectorsInCollection(collectionId);
|
||||
// 聊天级库按 ${chatId}_${taskId} 命名空间计数（getKbCollectionId 统一处理）
|
||||
return await countVectorsInCollection(getKbCollectionId(base, scope));
|
||||
|
||||
} else {
|
||||
if (settings.retrieval.independentChatMemoryEnabled) {
|
||||
const chatId = getChatId();
|
||||
if (!chatId) return 0;
|
||||
const totalCount = await countVectorsInCollection(chatId);
|
||||
console.log(`[翰林院-日志] 独立聊天记忆模式开启，聊天 ${chatId} 的向量总数: ${totalCount}`);
|
||||
return totalCount;
|
||||
}
|
||||
// 总数统计与查询侧保持同一可见性规则：
|
||||
// 独立模式 → 本地知识库 + 当前聊天的聊天级库 + 全局库
|
||||
// 统一模式 → 全部本地库（含聊天级）+ 全局库 + legacy 宝库
|
||||
const independent = settings.retrieval.independentChatMemoryEnabled;
|
||||
const chatId = independent ? getChatId() : null;
|
||||
console.log(`[翰林院-日志] 开始获取${independent ? '当前聊天可见的' : '所有'}知识库向量总数...`);
|
||||
|
||||
console.log('[翰林院-日志] 开始获取所有知识库的向量总数...');
|
||||
const localBases = Object.values(getLocalKnowledgeBases());
|
||||
const localBases = Object.values(getLocalKnowledgeBases())
|
||||
.filter(base => !independent || !base.chatId || base.chatId === chatId);
|
||||
const globalBases = Object.values(getGlobalKnowledgeBases());
|
||||
|
||||
const countPromises = [];
|
||||
|
||||
localBases.forEach(base => {
|
||||
const collectionId = `${charId}_${base.id}`;
|
||||
countPromises.push(countVectorsInCollection(collectionId));
|
||||
countPromises.push(countVectorsInCollection(getKbCollectionId(base, 'local')));
|
||||
});
|
||||
|
||||
globalBases.forEach(base => {
|
||||
const ownerId = base.owner || GLOBAL_SCOPE_ID;
|
||||
const collectionId = `${ownerId}_${base.id}`;
|
||||
countPromises.push(countVectorsInCollection(collectionId));
|
||||
countPromises.push(countVectorsInCollection(getKbCollectionId(base, 'global')));
|
||||
});
|
||||
|
||||
const legacyCollectionId = await getDynamicCollectionId();
|
||||
countPromises.push(countVectorsInCollection(legacyCollectionId));
|
||||
if (!independent) {
|
||||
const legacyCollectionId = await getDynamicCollectionId();
|
||||
countPromises.push(countVectorsInCollection(legacyCollectionId));
|
||||
}
|
||||
|
||||
const counts = await Promise.all(countPromises);
|
||||
const totalCount = counts.reduce((total, count) => total + count, 0);
|
||||
|
||||
console.log(`[翰林院-日志] 所有知识库统计完成，总向量数: ${totalCount}`);
|
||||
|
||||
console.log(`[翰林院-日志] 知识库统计完成，总向量数: ${totalCount}`);
|
||||
return totalCount;
|
||||
}
|
||||
}
|
||||
@@ -1202,20 +1310,23 @@ async function processCondensation(messages, logCallback = () => {}, range = nul
|
||||
kbName = `聊天记录: ${timestamp}`;
|
||||
}
|
||||
|
||||
const existingKbs = Object.values(getLocalKnowledgeBases());
|
||||
const foundKb = existingKbs.find(kb => kb.name === kbName);
|
||||
// 独立聊天记忆模式下凝识结果按聊天分桶，与 ingestTextToHanlinyuan 的语义一致
|
||||
const independentChatId = settings.retrieval.independentChatMemoryEnabled ? getChatId() : null;
|
||||
|
||||
const existingKbs = Object.values(getLocalKnowledgeBases());
|
||||
const foundKb = existingKbs.find(kb => kb.name === kbName && (kb.chatId ?? null) === independentChatId);
|
||||
|
||||
if (foundKb) {
|
||||
taskId = foundKb.id;
|
||||
logCallback(`[翰林院-核心] 检测到同名知识库 "${kbName}"，将数据合并入库。`, 'info');
|
||||
} else {
|
||||
logCallback(`[翰林院-核心] 准备为任务 "${kbName}" 创建专属知识库...`, 'info');
|
||||
const newKb = addKnowledgeBase(kbName, 'chat_history');
|
||||
const newKb = addKnowledgeBase(kbName, 'chat_history', independentChatId);
|
||||
taskId = newKb.id;
|
||||
}
|
||||
|
||||
|
||||
const charId = getCharacterStableId();
|
||||
const collectionId = `${charId}_${taskId}`;
|
||||
const collectionId = independentChatId ? `${independentChatId}_${taskId}` : `${charId}_${taskId}`;
|
||||
logCallback(`[翰林院-核心] 凝识任务已锁定知识库: ${kbName} (集合ID: ${collectionId})`, 'success');
|
||||
|
||||
const allChunks = [];
|
||||
@@ -1478,18 +1589,102 @@ async function rerankResults(allResults, queryText, settings) {
|
||||
finalScoredResults.sort((a, b) => (b.final_score || 0) - (a.final_score || 0));
|
||||
console.log('[翰林院-Rerank] 元数据加权排序完成。');
|
||||
|
||||
let finalResults = finalScoredResults;
|
||||
// 先按相关度截断 top_n，再做时序排序——顺序反了会让"时序最早"而非"最相关"
|
||||
// 的块占据名额（超级排序把最旧楼层排最前，slice 会扔掉高相关的靠后结果）
|
||||
let finalResults = finalScoredResults.slice(0, settings.rerank.top_n);
|
||||
if (settings.rerank.superSortEnabled) {
|
||||
finalResults = superSort(finalScoredResults);
|
||||
finalResults = superSort(finalResults);
|
||||
}
|
||||
|
||||
|
||||
return {
|
||||
results: finalResults.slice(0, settings.rerank.top_n),
|
||||
results: finalResults,
|
||||
reranked: rerankedSuccessfully
|
||||
};
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* 从"第十二章"/"第3卷"/"4"等字符串中解析序数，用于注入前的时序排序。
|
||||
* 支持阿拉伯数字与常见中文数字（至万级）；解析失败返回 MAX_SAFE_INTEGER（排最后）。
|
||||
*/
|
||||
function _parseOrdinal(value) {
|
||||
if (typeof value === 'number') return value;
|
||||
if (!value) return Number.MAX_SAFE_INTEGER;
|
||||
const str = String(value);
|
||||
const arabic = str.match(/\d+/);
|
||||
if (arabic) return parseInt(arabic[0], 10);
|
||||
|
||||
const cnDigit = { 零: 0, 一: 1, 二: 2, 两: 2, 三: 3, 四: 4, 五: 5, 六: 6, 七: 7, 八: 8, 九: 9 };
|
||||
const m = str.match(/[零一二两三四五六七八九十百千万]+/);
|
||||
if (!m) return Number.MAX_SAFE_INTEGER;
|
||||
let total = 0, current = 0;
|
||||
for (const ch of m[0]) {
|
||||
if (cnDigit[ch] !== undefined) {
|
||||
current = cnDigit[ch];
|
||||
} else if (ch === '十') {
|
||||
total += (current || 1) * 10;
|
||||
current = 0;
|
||||
} else if (ch === '百') {
|
||||
total += (current || 1) * 100;
|
||||
current = 0;
|
||||
} else if (ch === '千') {
|
||||
total += (current || 1) * 1000;
|
||||
current = 0;
|
||||
} else if (ch === '万') {
|
||||
total = (total + current) * 10000;
|
||||
current = 0;
|
||||
}
|
||||
}
|
||||
return total + current;
|
||||
}
|
||||
|
||||
/**
|
||||
* 注入前的组内时序重排 + 断层提示。
|
||||
*
|
||||
* rerank/相似度负责"选哪些块"，本函数负责"按什么顺序呈现"：
|
||||
* - chat_history 按楼层+部分升序；相邻块楼层跳跃时插入断层提示行，
|
||||
* 避免 LLM 把"不打不相识"和"关系亲密"两个远隔的片段读成连续剧情
|
||||
* - novel 按卷/章/节序数升序（中文数字章节号可解析）
|
||||
* - lorebook / manual 按来源聚合 + part 升序，碎块归位
|
||||
* 元数据缺失的块排在末尾、保持彼此原有顺序（sort 稳定性）。
|
||||
*/
|
||||
function _composeInjectionText(source, results) {
|
||||
const sorted = [...results];
|
||||
const ord = (v) => (Number.isFinite(v) ? v : Number.MAX_SAFE_INTEGER);
|
||||
|
||||
if (source === 'chat_history') {
|
||||
sorted.sort((a, b) =>
|
||||
ord(a.metadata?.floor) - ord(b.metadata?.floor)
|
||||
|| (a.metadata?.part ?? 0) - (b.metadata?.part ?? 0));
|
||||
|
||||
const parts = [];
|
||||
let prevFloor = null;
|
||||
for (const r of sorted) {
|
||||
const floor = r.metadata?.floor;
|
||||
if (prevFloor !== null && Number.isFinite(floor) && floor - prevFloor > 1) {
|
||||
parts.push(`〔提示：以下内容与上文相隔约 ${floor - prevFloor} 楼，期间的剧情未被检索到，两段内容并非连续发生〕`);
|
||||
}
|
||||
parts.push(r.text);
|
||||
if (Number.isFinite(floor)) prevFloor = floor;
|
||||
}
|
||||
return parts.join('\n\n');
|
||||
}
|
||||
|
||||
if (source === 'novel') {
|
||||
sorted.sort((a, b) =>
|
||||
_parseOrdinal(a.metadata?.volume) - _parseOrdinal(b.metadata?.volume)
|
||||
|| _parseOrdinal(a.metadata?.chapter) - _parseOrdinal(b.metadata?.chapter)
|
||||
|| _parseOrdinal(a.metadata?.section) - _parseOrdinal(b.metadata?.section));
|
||||
return sorted.map(r => r.text).join('\n\n');
|
||||
}
|
||||
|
||||
// lorebook / manual：同源聚合 + part 升序
|
||||
sorted.sort((a, b) =>
|
||||
String(a.metadata?.sourceName ?? '').localeCompare(String(b.metadata?.sourceName ?? ''), 'zh')
|
||||
|| (a.metadata?.part ?? 0) - (b.metadata?.part ?? 0));
|
||||
return sorted.map(r => r.text).join('\n\n');
|
||||
}
|
||||
|
||||
async function rearrangeChat(chat, contextSize, abort, type) {
|
||||
const injectionKeys = {
|
||||
novel: 'HANLINYUAN_RAG_NOVEL',
|
||||
@@ -1704,7 +1899,8 @@ async function rearrangeChat(chat, contextSize, abort, type) {
|
||||
continue;
|
||||
}
|
||||
|
||||
const formattedText = results.map(r => r.text).join('\n\n');
|
||||
// 组内按时序重排 + 断层提示（rerank 决定选哪些块，时序决定呈现顺序）
|
||||
const formattedText = _composeInjectionText(source, results);
|
||||
const placeholder = `{{${source.replace('_history', '')}_text}}`;
|
||||
let injectionContent = injectionSettings.template.replace(placeholder, formattedText);
|
||||
|
||||
@@ -1751,6 +1947,13 @@ async function moveKnowledgeBase(taskId, fromScope) {
|
||||
return;
|
||||
}
|
||||
|
||||
// 聊天级库（独立聊天记忆产物）专属于单个聊天，移到全局会让所有角色
|
||||
// 检索到某个特定聊天的记忆，语义矛盾，禁止
|
||||
if (kbData.chatId && toScope === 'global') {
|
||||
toastr.warning(`知识库【${kbData.name}】是聊天专属记忆，不能移动到全局。`);
|
||||
return;
|
||||
}
|
||||
|
||||
if (fromScope === 'local' && toScope === 'global' && !kbData.owner) {
|
||||
console.log(`[翰林院-配置] 为旧版知识库 ${taskId} 补充所有者ID: ${charId}`);
|
||||
kbData.owner = charId;
|
||||
|
||||
Reference in New Issue
Block a user