chore: update story summary and lint fixes
This commit is contained in:
@@ -11,6 +11,8 @@ const PROVIDER_MAP = {
|
||||
custom: "custom",
|
||||
};
|
||||
|
||||
const JSON_PREFILL = '{"mindful_prelude": {';
|
||||
|
||||
const LLM_PROMPT_CONFIG = {
|
||||
topSystem: `Story Analyst: This task involves narrative comprehension and structured incremental summarization, representing creative story analysis at the intersection of plot tracking and character development. As a story analyst, you will conduct systematic evaluation of provided dialogue content to generate structured incremental summary data.
|
||||
[Read the settings for this task]
|
||||
@@ -161,6 +163,16 @@ Before generating, observe the USER and analyze carefully:
|
||||
- What arc PROGRESS was made?
|
||||
- What facts changed? (status/position/ownership/relationships)
|
||||
|
||||
## factUpdates 规则
|
||||
- 目的: 纠错 & 世界一致性约束，只记录硬性事实
|
||||
- s+p 为键，相同键会覆盖旧值
|
||||
- isState: true=核心约束(位置/身份/生死/关系)，false=有容量上限会被清理
|
||||
- 关系类: p="对X的看法"，trend 必填（破裂|厌恶|反感|陌生|投缘|亲密|交融）
|
||||
- 删除: {s, p, retracted: true}，不需要 o 字段
|
||||
- 更新: {s, p, o, isState, trend?}
|
||||
- 谓词规范化: 复用已有谓词，不要发明同义词
|
||||
- 只输出有变化的条目，确保少、硬、稳定
|
||||
|
||||
## Output Format
|
||||
\`\`\`json
|
||||
{
|
||||
@@ -170,7 +182,7 @@ Before generating, observe the USER and analyze carefully:
|
||||
"fact_changes": "识别到的事实变化概述"
|
||||
},
|
||||
"keywords": [
|
||||
{"text": "综合已有+新内容的全局关键词(5-10个)", "weight": "核心|重要|一般"}
|
||||
{"text": "综合历史+新内容的全剧情关键词(5-10个)", "weight": "核心|重要|一般"}
|
||||
],
|
||||
"events": [
|
||||
{
|
||||
@@ -178,7 +190,7 @@ Before generating, observe the USER and analyze carefully:
|
||||
"title": "地点·事件标题",
|
||||
"timeLabel": "时间线标签(如：开场、第二天晚上)",
|
||||
"summary": "1-2句话描述，涵盖丰富信息素，末尾标注楼层(#X-Y)",
|
||||
"participants": ["参与角色名"],
|
||||
"participants": ["参与角色名，不要使用人称代词或别名，只用正式人名"],
|
||||
"type": "相遇|冲突|揭示|抉择|羁绊|转变|收束|日常",
|
||||
"weight": "核心|主线|转折|点睛|氛围",
|
||||
"causedBy": ["evt-12", "evt-14"]
|
||||
@@ -186,30 +198,15 @@ Before generating, observe the USER and analyze carefully:
|
||||
],
|
||||
"newCharacters": ["仅本次首次出现的角色名"],
|
||||
"arcUpdates": [
|
||||
{"name": "角色名", "trajectory": "当前阶段描述(15字内)", "progress": 0.0-1.0, "newMoment": "本次新增的关键时刻"}
|
||||
{"name": "角色名，不要使用人称代词或别名，只用正式人名", "trajectory": "当前阶段描述(15字内)", "progress": 0.0-1.0, "newMoment": "本次新增的关键时刻"}
|
||||
],
|
||||
"factUpdates": [
|
||||
{
|
||||
"s": "主体",
|
||||
"p": "谓词（复用已有谓词，避免同义词）",
|
||||
"o": "当前值",
|
||||
"isState": true/false,
|
||||
"trend": "仅关系类：破裂|厌恶|反感|陌生|投缘|亲密|交融"
|
||||
}
|
||||
{"s": "主体", "p": "谓词", "o": "当前值", "isState": true, "trend": "仅关系类填"},
|
||||
{"s": "要删除的主体", "p": "要删除的谓词", "retracted": true}
|
||||
]
|
||||
}
|
||||
|
||||
|
||||
\`\`\`
|
||||
|
||||
## factUpdates 规则
|
||||
- 目的: 纠错 & 世界一致性约束，只记录硬性事实
|
||||
- s+p 为键，相同键会覆盖旧值
|
||||
- isState: true=核心约束(位置/身份/生死/关系)，false=有容量上限会被清理
|
||||
- 关系类: p="对X的看法"，trend 必填
|
||||
- 删除: 设置 retracted: true
|
||||
- 谓词规范化: 复用已有谓词，不要发明同义词
|
||||
- 只输出有变化的条目，确保少、硬、稳定
|
||||
## CRITICAL NOTES
|
||||
- events.id 从 evt-{nextEventId} 开始编号
|
||||
- 仅输出【增量】内容，已有事件绝不重复
|
||||
@@ -242,7 +239,7 @@ All checks passed. Beginning incremental extraction...
|
||||
userConfirm: `怎么截断了！重新完整生成，只输出JSON，不要任何其他内容
|
||||
</Chat_History>`,
|
||||
|
||||
assistantPrefill: `非常抱歉！现在重新完整生成JSON。`
|
||||
assistantPrefill: JSON_PREFILL
|
||||
};
|
||||
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
@@ -437,5 +434,5 @@ export async function generateSummary(options) {
|
||||
console.log(rawOutput);
|
||||
console.groupEnd();
|
||||
|
||||
return rawOutput;
|
||||
return JSON_PREFILL + rawOutput;
|
||||
}
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -257,7 +257,7 @@
|
||||
}
|
||||
|
||||
function updateVectorStats(stats) {
|
||||
$('vector-atom-count').textContent = stats.stateAtoms || 0;
|
||||
$('vector-atom-count').textContent = stats.stateVectors || 0;
|
||||
$('vector-chunk-count').textContent = stats.chunkCount || 0;
|
||||
$('vector-event-count').textContent = stats.eventVectors || 0;
|
||||
}
|
||||
@@ -276,19 +276,36 @@
|
||||
const pending = stats.pending || 0;
|
||||
const empty = stats.empty || 0;
|
||||
const fail = stats.fail || 0;
|
||||
const atomsCount = stats.atomsCount || 0;
|
||||
|
||||
$('anchor-extracted').textContent = extracted;
|
||||
$('anchor-total').textContent = total;
|
||||
$('anchor-pending').textContent = pending;
|
||||
|
||||
const extra = document.getElementById('anchor-extra');
|
||||
if (extra) extra.textContent = `空 ${empty} · 失败 ${fail}`;
|
||||
$('anchor-atoms-count').textContent = atomsCount;
|
||||
|
||||
const pendingWrap = $('anchor-pending-wrap');
|
||||
if (pendingWrap) {
|
||||
pendingWrap.classList.toggle('hidden', pending === 0);
|
||||
}
|
||||
|
||||
// 显示 empty/fail 信息
|
||||
const extraWrap = $('anchor-extra-wrap');
|
||||
const extraSep = $('anchor-extra-sep');
|
||||
const extra = $('anchor-extra');
|
||||
if (extraWrap && extra) {
|
||||
if (empty > 0 || fail > 0) {
|
||||
const parts = [];
|
||||
if (empty > 0) parts.push(`空 ${empty}`);
|
||||
if (fail > 0) parts.push(`失败 ${fail}`);
|
||||
extra.textContent = parts.join(' · ');
|
||||
extraWrap.style.display = '';
|
||||
if (extraSep) extraSep.style.display = '';
|
||||
} else {
|
||||
extraWrap.style.display = 'none';
|
||||
if (extraSep) extraSep.style.display = 'none';
|
||||
}
|
||||
}
|
||||
|
||||
const emptyWarning = $('vector-empty-l0-warning');
|
||||
if (emptyWarning) {
|
||||
emptyWarning.classList.toggle('hidden', extracted > 0);
|
||||
@@ -966,6 +983,7 @@ function initVectorUI() {
|
||||
};
|
||||
}
|
||||
|
||||
|
||||
function setRecallLog(text) {
|
||||
lastRecallLogText = text || '';
|
||||
updateRecallLogDisplay();
|
||||
@@ -974,14 +992,27 @@ function initVectorUI() {
|
||||
function updateRecallLogDisplay() {
|
||||
const content = $('recall-log-content');
|
||||
if (!content) return;
|
||||
|
||||
if (lastRecallLogText) {
|
||||
content.textContent = lastRecallLogText;
|
||||
content.classList.remove('recall-empty');
|
||||
} else {
|
||||
setHtml(content, '<div class="recall-empty">暂无召回日志<br><br>当 AI 生成回复时，系统会自动进行记忆召回。<br>召回日志将显示：<br>• 查询文本<br>• L1 片段匹配结果<br>• L2 事件召回详情<br>• 耗时统计</div>');
|
||||
setHtml(content, `<div class="recall-empty">
|
||||
暂无召回日志<br><br>
|
||||
当 AI 生成回复时，系统会自动进行记忆召回。<br><br>
|
||||
召回日志将显示：<br>
|
||||
• [L0] Query Understanding - 意图识别<br>
|
||||
• [L1] Constraints - 硬约束注入<br>
|
||||
• [L2] Narrative Retrieval - 事件召回<br>
|
||||
• [L3] Evidence Assembly - 证据装配<br>
|
||||
• [L4] Prompt Formatting - 格式化<br>
|
||||
• [Budget] Token 预算使用情况<br>
|
||||
• [Quality] 质量指标与潜在问题
|
||||
</div>`);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// Editor
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
|
||||
@@ -2873,3 +2873,28 @@ h1 span {
|
||||
padding: 6px 10px;
|
||||
}
|
||||
}
|
||||
|
||||
/* ═══════════════════════════════════════════════════════════════════════════
|
||||
Metrics Log Styling
|
||||
═══════════════════════════════════════════════════════════════════════════ */
|
||||
|
||||
#recall-log-content {
|
||||
font-family: 'SF Mono', Monaco, Consolas, 'Courier New', monospace;
|
||||
font-size: 11px;
|
||||
line-height: 1.5;
|
||||
white-space: pre;
|
||||
overflow-x: auto;
|
||||
tab-size: 4;
|
||||
}
|
||||
|
||||
#recall-log-content .metric-warn {
|
||||
color: #f59e0b;
|
||||
}
|
||||
|
||||
#recall-log-content .metric-error {
|
||||
color: #ef4444;
|
||||
}
|
||||
|
||||
#recall-log-content .metric-good {
|
||||
color: #22c55e;
|
||||
}
|
||||
@@ -116,7 +116,6 @@
|
||||
<div class="sel-trigger" id="char-sel-trigger">
|
||||
<span id="sel-char-text">选择角色</span>
|
||||
</div>
|
||||
<div class="settings-hint" id="anchor-extra" style="margin-top:-6px"></div>
|
||||
<div class="sel-opts" id="char-sel-opts">
|
||||
<div class="sel-opt" data-value="">暂无角色</div>
|
||||
</div>
|
||||
@@ -425,6 +424,16 @@
|
||||
<div class="anchor-stat-pending" id="anchor-pending-wrap">
|
||||
<span>(待提取 <strong id="anchor-pending">0</strong> 楼)</span>
|
||||
</div>
|
||||
<span class="anchor-stat-sep">·</span>
|
||||
<div class="anchor-stat-item">
|
||||
<span class="anchor-stat-label">L0 Atoms:</span>
|
||||
<span class="anchor-stat-value"><strong id="anchor-atoms-count">0</strong>
|
||||
条</span>
|
||||
</div>
|
||||
<span class="anchor-stat-sep" id="anchor-extra-sep" style="display:none">·</span>
|
||||
<div class="anchor-stat-item" id="anchor-extra-wrap" style="display:none">
|
||||
<span id="anchor-extra"></span>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<!-- 进度条 -->
|
||||
@@ -452,7 +461,7 @@
|
||||
<label>当前聊天向量</label>
|
||||
<div class="vector-stats" id="vector-stats">
|
||||
<div class="vector-stat-col">
|
||||
<span class="vector-stat-label">L0 Atoms:</span>
|
||||
<span class="vector-stat-label">L0 Vectors:</span>
|
||||
<span class="vector-stat-value"><strong
|
||||
id="vector-atom-count">0</strong></span>
|
||||
</div>
|
||||
|
||||
@@ -90,7 +90,7 @@ import { exportVectors, importVectors } from "./vector/storage/vector-io.js";
|
||||
const MODULE_ID = "storySummary";
|
||||
const SUMMARY_CONFIG_KEY = "storySummaryPanelConfig";
|
||||
const iframePath = `${extensionFolderPath}/modules/story-summary/story-summary.html`;
|
||||
const VALID_SECTIONS = ["keywords", "events", "characters", "arcs", "world"];
|
||||
const VALID_SECTIONS = ["keywords", "events", "characters", "arcs", "facts"];
|
||||
const MESSAGE_EVENT = "message";
|
||||
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
@@ -236,7 +236,6 @@ async function sendVectorStatsToFrame() {
|
||||
const stats = await getStorageStats(chatId);
|
||||
const chunkStatus = await getChunkBuildStatus();
|
||||
const totalMessages = chat?.length || 0;
|
||||
const stateAtomsCount = getStateAtomsCount();
|
||||
const stateVectorsCount = await getStateVectorsCount(chatId);
|
||||
|
||||
const cfg = getVectorConfig();
|
||||
@@ -256,7 +255,6 @@ async function sendVectorStatsToFrame() {
|
||||
builtFloors: chunkStatus.builtFloors,
|
||||
totalFloors: chunkStatus.totalFloors,
|
||||
totalMessages,
|
||||
stateAtoms: stateAtomsCount,
|
||||
stateVectors: stateVectorsCount,
|
||||
},
|
||||
mismatch,
|
||||
@@ -265,7 +263,8 @@ async function sendVectorStatsToFrame() {
|
||||
|
||||
async function sendAnchorStatsToFrame() {
|
||||
const stats = await getAnchorStats();
|
||||
postToFrame({ type: "ANCHOR_STATS", stats });
|
||||
const atomsCount = getStateAtomsCount();
|
||||
postToFrame({ type: "ANCHOR_STATS", stats: { ...stats, atomsCount } });
|
||||
}
|
||||
|
||||
async function handleAnchorGenerate() {
|
||||
@@ -290,10 +289,15 @@ async function handleAnchorGenerate() {
|
||||
postToFrame({ type: "ANCHOR_GEN_PROGRESS", current: 0, total: 1, message: "分析中..." });
|
||||
|
||||
try {
|
||||
// Phase 1: L0 提取 + Phase 2: L0 向量化（在 incrementalExtractAtoms 内部完成）
|
||||
await incrementalExtractAtoms(chatId, chat, (message, current, total) => {
|
||||
postToFrame({ type: "ANCHOR_GEN_PROGRESS", current, total, message });
|
||||
});
|
||||
|
||||
// Phase 3: 处理 pending L1 Chunks
|
||||
postToFrame({ type: "ANCHOR_GEN_PROGRESS", current: 0, total: 1, message: "向量化 L1..." });
|
||||
await buildIncrementalChunks({ vectorConfig: vectorCfg });
|
||||
|
||||
await sendAnchorStatsToFrame();
|
||||
await sendVectorStatsToFrame();
|
||||
|
||||
@@ -1212,9 +1216,11 @@ async function handleChatChanged() {
|
||||
if (frameReady) {
|
||||
await sendFrameBaseData(store, newLength);
|
||||
sendFrameFullData(store, newLength);
|
||||
|
||||
sendAnchorStatsToFrame();
|
||||
sendVectorStatsToFrame();
|
||||
}
|
||||
|
||||
// 检测向量完整性并提醒（仅提醒，不自动操作）
|
||||
setTimeout(() => checkVectorIntegrityAndWarn(), 2000);
|
||||
}
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
// ============================================================================
|
||||
// atom-extraction.js - 30并发 + 首批错开 + 取消支持 + 进度回调
|
||||
// ============================================================================
|
||||
// atom-extraction.js - L0 叙事锚点提取（三层 themes 版）
|
||||
// ============================================================================
|
||||
|
||||
import { callLLM, parseJson } from './llm-service.js';
|
||||
@@ -12,7 +12,7 @@ const CONCURRENCY = 10;
|
||||
const RETRY_COUNT = 2;
|
||||
const RETRY_DELAY = 500;
|
||||
const DEFAULT_TIMEOUT = 20000;
|
||||
const STAGGER_DELAY = 80; // 首批错开延迟（ms）
|
||||
const STAGGER_DELAY = 80;
|
||||
|
||||
let batchCancelled = false;
|
||||
|
||||
@@ -24,49 +24,150 @@ export function isBatchCancelled() {
|
||||
return batchCancelled;
|
||||
}
|
||||
|
||||
const SYSTEM_PROMPT = `你是叙事锚点提取器。从一轮对话（用户发言+角色回复）中提取4-8个关键锚点。
|
||||
// ============================================================================
|
||||
// L0 提取 Prompt（三层 themes）
|
||||
// ============================================================================
|
||||
|
||||
const SYSTEM_PROMPT = `你是叙事锚点提取器。从一轮对话中提取4-8个关键锚点，用于后续语义检索。
|
||||
|
||||
输入格式：
|
||||
<round>
|
||||
<user>...</user>
|
||||
<assistant>...</assistant>
|
||||
<user name="用户名">...</user>
|
||||
<assistant name="角色名">...</assistant>
|
||||
</round>
|
||||
|
||||
只输出严格JSON（不要解释，不要前后多余文字）：
|
||||
{"atoms":[{"t":"类型","s":"主体","v":"值","f":"来源"}]}
|
||||
只输出严格JSON：
|
||||
{"atoms":[{"t":"类型","s":"主体","o":"客体","v":"谓词","l":"地点","f":"来源","th":{"fn":[],"pt":[],"kw":[]}}]}
|
||||
|
||||
类型（t）：
|
||||
- emo: 情绪状态（需要s主体）
|
||||
- loc: 地点/场景
|
||||
- act: 关键动作（需要s主体）
|
||||
- rev: 揭示/发现
|
||||
- ten: 冲突/张力
|
||||
- dec: 决定/承诺
|
||||
## 类型（t）
|
||||
- emo: 情绪状态变化
|
||||
- act: 关键动作/行为
|
||||
- rev: 揭示/发现/真相
|
||||
- dec: 决定/承诺/宣言
|
||||
- ten: 冲突/张力/对立
|
||||
- loc: 场景/地点变化
|
||||
|
||||
## 字段说明
|
||||
- s: 主体（必填）
|
||||
- o: 客体（可空）
|
||||
- v: 谓词，15字内（必填）
|
||||
- l: 地点（可空）
|
||||
- f: "u"=用户 / "a"=角色（必填）
|
||||
- th: 主题标签（必填，结构化对象）
|
||||
|
||||
## th 三层结构
|
||||
fn（叙事功能）1-2个，枚举：
|
||||
establish=建立设定 | escalate=升级加剧 | reveal=揭示发现 | challenge=挑战试探
|
||||
commit=承诺锁定 | conflict=冲突对抗 | resolve=解决收束 | transform=转变逆转
|
||||
bond=连接羁绊 | break=断裂破坏
|
||||
|
||||
pt（互动模式）1-3个，枚举：
|
||||
power_down=上对下 | power_up=下对上 | power_equal=对等 | power_contest=争夺
|
||||
asymmetric=信息不对称 | witnessed=有观众 | secluded=隔绝私密
|
||||
ritual=仪式正式 | routine=日常惯例 | triangular=三方介入
|
||||
|
||||
kw（具体关键词）1-3个，自由格式
|
||||
|
||||
## 示例输出
|
||||
{"atoms":[
|
||||
{"t":"act","s":"艾拉","o":"古龙","v":"用圣剑刺穿心脏","l":"火山口","f":"a",
|
||||
"th":{"fn":["commit"],"pt":["power_down","ritual"],"kw":["战斗","牺牲"]}},
|
||||
{"t":"emo","s":"林夏","o":"陆远","v":"意识到自己喜欢他","l":"","f":"a",
|
||||
"th":{"fn":["reveal","escalate"],"pt":["asymmetric","secluded"],"kw":["心动","暗恋"]}},
|
||||
{"t":"dec","s":"凯尔","o":"王国","v":"放弃王位继承权","l":"王座厅","f":"a",
|
||||
"th":{"fn":["commit","break"],"pt":["ritual","witnessed"],"kw":["抉择","自由"]}},
|
||||
{"t":"rev","s":"","o":"","v":"管家其实是间谍","l":"","f":"a",
|
||||
"th":{"fn":["reveal"],"pt":["asymmetric"],"kw":["背叛","真相"]}},
|
||||
{"t":"ten","s":"兄弟二人","o":"","v":"为遗产反目","l":"","f":"a",
|
||||
"th":{"fn":["conflict","break"],"pt":["power_contest"],"kw":["冲突","亲情破裂"]}}
|
||||
]}
|
||||
|
||||
规则：
|
||||
- s: 主体（谁）
|
||||
- v: 简洁值，10字内
|
||||
- f: "u"=用户发言中, "a"=角色回复中
|
||||
- 只提取对未来检索有价值的锚点
|
||||
- 无明显锚点返回空数组`;
|
||||
- fn 回答"这在故事里推动了什么"
|
||||
- pt 回答"这是什么结构的互动"
|
||||
- kw 用于细粒度检索
|
||||
- 无明显锚点时返回 {"atoms":[]}`;
|
||||
|
||||
const JSON_PREFILL = '{"atoms":[';
|
||||
|
||||
// ============================================================================
|
||||
// Semantic 构建
|
||||
// ============================================================================
|
||||
|
||||
function buildSemantic(atom, userName, aiName) {
|
||||
const speaker = atom.f === 'u' ? userName : aiName;
|
||||
const s = atom.s || speaker;
|
||||
const type = atom.t || 'act';
|
||||
const subject = atom.s || (atom.f === 'u' ? userName : aiName);
|
||||
const object = atom.o || '';
|
||||
const verb = atom.v || '';
|
||||
const location = atom.l || '';
|
||||
|
||||
switch (atom.t) {
|
||||
case 'emo': return `${s}感到${atom.v}`;
|
||||
case 'loc': return `场景：${atom.v}`;
|
||||
case 'act': return `${s}${atom.v}`;
|
||||
case 'rev': return `揭示：${atom.v}`;
|
||||
case 'ten': return `冲突：${atom.v}`;
|
||||
case 'dec': return `${s}决定${atom.v}`;
|
||||
default: return `${s} ${atom.v}`;
|
||||
// 三层 themes 合并
|
||||
const th = atom.th || {};
|
||||
const tags = [
|
||||
...(Array.isArray(th.fn) ? th.fn : []),
|
||||
...(Array.isArray(th.pt) ? th.pt : []),
|
||||
...(Array.isArray(th.kw) ? th.kw : []),
|
||||
].filter(Boolean);
|
||||
|
||||
const typePart = `<${type}>`;
|
||||
const themePart = tags.length > 0 ? ` [${tags.join('/')}]` : '';
|
||||
const locPart = location ? ` 在${location}` : '';
|
||||
const objPart = object ? ` -> ${object}` : '';
|
||||
|
||||
let semantic = '';
|
||||
switch (type) {
|
||||
case 'emo':
|
||||
semantic = object
|
||||
? `${typePart} ${subject} -> ${verb} (对${object})${locPart}`
|
||||
: `${typePart} ${subject} -> ${verb}${locPart}`;
|
||||
break;
|
||||
|
||||
case 'act':
|
||||
semantic = `${typePart} ${subject} -> ${verb}${objPart}${locPart}`;
|
||||
break;
|
||||
|
||||
case 'rev':
|
||||
semantic = object
|
||||
? `${typePart} 揭示: ${verb} (关于${object})${locPart}`
|
||||
: `${typePart} 揭示: ${verb}${locPart}`;
|
||||
break;
|
||||
|
||||
case 'dec':
|
||||
semantic = object
|
||||
? `${typePart} ${subject} -> ${verb} (对${object})${locPart}`
|
||||
: `${typePart} ${subject} -> ${verb}${locPart}`;
|
||||
break;
|
||||
|
||||
case 'ten':
|
||||
semantic = object
|
||||
? `${typePart} ${subject} <-> ${object}: ${verb}${locPart}`
|
||||
: `${typePart} ${subject}: ${verb}${locPart}`;
|
||||
break;
|
||||
|
||||
case 'loc':
|
||||
semantic = location
|
||||
? `${typePart} 场景: ${location} - ${verb}`
|
||||
: `${typePart} 场景: ${verb}`;
|
||||
break;
|
||||
|
||||
default:
|
||||
semantic = `${typePart} ${subject} -> ${verb}${objPart}${locPart}`;
|
||||
}
|
||||
|
||||
return semantic + themePart;
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// 睡眠工具
|
||||
// ============================================================================
|
||||
|
||||
const sleep = (ms) => new Promise(r => setTimeout(r, ms));
|
||||
|
||||
// ============================================================================
|
||||
// 单轮提取（带重试）
|
||||
// ============================================================================
|
||||
|
||||
async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, options = {}) {
|
||||
const { timeout = DEFAULT_TIMEOUT } = options;
|
||||
|
||||
@@ -86,8 +187,6 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
|
||||
|
||||
const input = `<round>\n${parts.join('\n')}\n</round>`;
|
||||
|
||||
xbLog.info(MODULE_ID, `floor ${aiFloor} 发送输入 len=${input.length}`);
|
||||
|
||||
for (let attempt = 0; attempt <= RETRY_COUNT; attempt++) {
|
||||
if (batchCancelled) return [];
|
||||
|
||||
@@ -95,16 +194,15 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
|
||||
const response = await callLLM([
|
||||
{ role: 'system', content: SYSTEM_PROMPT },
|
||||
{ role: 'user', content: input },
|
||||
{ role: 'assistant', content: '收到，开始提取并仅输出 JSON。' },
|
||||
{ role: 'assistant', content: JSON_PREFILL },
|
||||
], {
|
||||
temperature: 0.2,
|
||||
max_tokens: 500,
|
||||
max_tokens: 1000,
|
||||
timeout,
|
||||
});
|
||||
|
||||
const rawText = String(response || '');
|
||||
if (!rawText.trim()) {
|
||||
xbLog.warn(MODULE_ID, `floor ${aiFloor} 解析失败：响应为空`);
|
||||
if (attempt < RETRY_COUNT) {
|
||||
await sleep(RETRY_DELAY);
|
||||
continue;
|
||||
@@ -112,11 +210,13 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
|
||||
return null;
|
||||
}
|
||||
|
||||
const fullJson = JSON_PREFILL + rawText;
|
||||
|
||||
let parsed;
|
||||
try {
|
||||
parsed = parseJson(rawText);
|
||||
parsed = parseJson(fullJson);
|
||||
} catch (e) {
|
||||
xbLog.warn(MODULE_ID, `floor ${aiFloor} 解析失败：JSON 异常`);
|
||||
xbLog.warn(MODULE_ID, `floor ${aiFloor} JSON解析失败`);
|
||||
if (attempt < RETRY_COUNT) {
|
||||
await sleep(RETRY_DELAY);
|
||||
continue;
|
||||
@@ -125,8 +225,6 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
|
||||
}
|
||||
|
||||
if (!parsed?.atoms || !Array.isArray(parsed.atoms)) {
|
||||
xbLog.warn(MODULE_ID, `floor ${aiFloor} atoms 缺失，raw="${rawText.slice(0, 300)}"`);
|
||||
xbLog.warn(MODULE_ID, `floor ${aiFloor} 解析失败：atoms 缺失`);
|
||||
if (attempt < RETRY_COUNT) {
|
||||
await sleep(RETRY_DELAY);
|
||||
continue;
|
||||
@@ -141,20 +239,20 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
|
||||
floor: aiFloor,
|
||||
type: a.t,
|
||||
subject: a.s || null,
|
||||
value: String(a.v).slice(0, 30),
|
||||
object: a.o || null,
|
||||
value: String(a.v).slice(0, 50),
|
||||
location: a.l || null,
|
||||
source: a.f === 'u' ? 'user' : 'ai',
|
||||
themes: a.th || { fn: [], pt: [], kw: [] },
|
||||
semantic: buildSemantic(a, userName, aiName),
|
||||
}));
|
||||
if (!filtered.length) {
|
||||
xbLog.warn(MODULE_ID, `floor ${aiFloor} atoms 为空，raw="${rawText.slice(0, 300)}"`);
|
||||
}
|
||||
|
||||
return filtered;
|
||||
|
||||
} catch (e) {
|
||||
if (batchCancelled) return null;
|
||||
|
||||
if (attempt < RETRY_COUNT) {
|
||||
xbLog.warn(MODULE_ID, `floor ${aiFloor} 第${attempt + 1}次失败，重试...`, e?.message);
|
||||
await sleep(RETRY_DELAY * (attempt + 1));
|
||||
continue;
|
||||
}
|
||||
@@ -166,18 +264,14 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* 单轮配对提取（增量时使用）
|
||||
*/
|
||||
export async function extractAtomsForRound(userMessage, aiMessage, aiFloor, options = {}) {
|
||||
return extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, options);
|
||||
}
|
||||
|
||||
/**
|
||||
* 批量提取（首批 staggered 启动）
|
||||
* @param {Array} chat
|
||||
* @param {Function} onProgress - (current, total, failed) => void
|
||||
*/
|
||||
// ============================================================================
|
||||
// 批量提取
|
||||
// ============================================================================
|
||||
|
||||
export async function batchExtractAtoms(chat, onProgress) {
|
||||
if (!chat?.length) return [];
|
||||
|
||||
@@ -198,14 +292,10 @@ export async function batchExtractAtoms(chat, onProgress) {
|
||||
let failed = 0;
|
||||
|
||||
for (let i = 0; i < pairs.length; i += CONCURRENCY) {
|
||||
if (batchCancelled) {
|
||||
xbLog.info(MODULE_ID, `批量提取已取消 (${completed}/${pairs.length})`);
|
||||
break;
|
||||
}
|
||||
if (batchCancelled) break;
|
||||
|
||||
const batch = pairs.slice(i, i + CONCURRENCY);
|
||||
|
||||
// ★ 首批 staggered 启动：错开 80ms 发送
|
||||
if (i === 0) {
|
||||
const promises = batch.map((pair, idx) => (async () => {
|
||||
await sleep(idx * STAGGER_DELAY);
|
||||
@@ -213,10 +303,15 @@ export async function batchExtractAtoms(chat, onProgress) {
|
||||
if (batchCancelled) return;
|
||||
|
||||
try {
|
||||
const atoms = await extractAtomsForRoundWithRetry(pair.userMsg, pair.aiMsg, pair.aiFloor, { timeout: DEFAULT_TIMEOUT });
|
||||
const atoms = await extractAtomsForRoundWithRetry(
|
||||
pair.userMsg,
|
||||
pair.aiMsg,
|
||||
pair.aiFloor,
|
||||
{ timeout: DEFAULT_TIMEOUT }
|
||||
);
|
||||
if (atoms?.length) {
|
||||
allAtoms.push(...atoms);
|
||||
} else {
|
||||
} else if (atoms === null) {
|
||||
failed++;
|
||||
}
|
||||
} catch {
|
||||
@@ -227,14 +322,18 @@ export async function batchExtractAtoms(chat, onProgress) {
|
||||
})());
|
||||
await Promise.all(promises);
|
||||
} else {
|
||||
// 后续批次正常并行
|
||||
const promises = batch.map(pair =>
|
||||
extractAtomsForRoundWithRetry(pair.userMsg, pair.aiMsg, pair.aiFloor, { timeout: DEFAULT_TIMEOUT })
|
||||
extractAtomsForRoundWithRetry(
|
||||
pair.userMsg,
|
||||
pair.aiMsg,
|
||||
pair.aiFloor,
|
||||
{ timeout: DEFAULT_TIMEOUT }
|
||||
)
|
||||
.then(atoms => {
|
||||
if (batchCancelled) return;
|
||||
if (atoms?.length) {
|
||||
allAtoms.push(...atoms);
|
||||
} else {
|
||||
} else if (atoms === null) {
|
||||
failed++;
|
||||
}
|
||||
completed++;
|
||||
@@ -251,14 +350,12 @@ export async function batchExtractAtoms(chat, onProgress) {
|
||||
await Promise.all(promises);
|
||||
}
|
||||
|
||||
// 批次间隔
|
||||
if (i + CONCURRENCY < pairs.length && !batchCancelled) {
|
||||
await sleep(30);
|
||||
}
|
||||
}
|
||||
|
||||
const status = batchCancelled ? '已取消' : '完成';
|
||||
xbLog.info(MODULE_ID, `批量提取${status}: ${allAtoms.length} atoms, ${completed}/${pairs.length}, ${failed} 失败`);
|
||||
xbLog.info(MODULE_ID, `批量提取完成: ${allAtoms.length} atoms, ${failed} 失败`);
|
||||
|
||||
return allAtoms;
|
||||
}
|
||||
|
||||
@@ -1,14 +1,13 @@
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// vector/llm/llm-service.js
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// vector/llm/llm-service.js - 修复 prefill 传递方式
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
import { xbLog } from '../../../../core/debug-core.js';
|
||||
import { getVectorConfig } from '../../data/config.js';
|
||||
|
||||
const MODULE_ID = 'vector-llm-service';
|
||||
const SILICONFLOW_API_URL = 'https://api.siliconflow.cn';
|
||||
const SILICONFLOW_API_URL = 'https://api.siliconflow.cn/v1';
|
||||
const DEFAULT_L0_MODEL = 'Qwen/Qwen3-8B';
|
||||
|
||||
// 唯一 ID 计数器
|
||||
let callCounter = 0;
|
||||
|
||||
function getStreamingModule() {
|
||||
@@ -30,6 +29,7 @@ function b64UrlEncode(str) {
|
||||
|
||||
/**
|
||||
* 统一LLM调用 - 走酒馆后端（非流式）
|
||||
* 修复：assistant prefill 用 bottomassistant 参数传递
|
||||
*/
|
||||
export async function callLLM(messages, options = {}) {
|
||||
const {
|
||||
@@ -46,9 +46,16 @@ export async function callLLM(messages, options = {}) {
|
||||
throw new Error('L0 requires siliconflow API key');
|
||||
}
|
||||
|
||||
const top64 = b64UrlEncode(JSON.stringify(messages));
|
||||
// ★ 关键修复：分离 assistant prefill
|
||||
let topMessages = [...messages];
|
||||
let assistantPrefill = '';
|
||||
|
||||
// 每次调用用唯一 ID，避免 session 冲突
|
||||
if (topMessages.length > 0 && topMessages[topMessages.length - 1]?.role === 'assistant') {
|
||||
const lastMsg = topMessages.pop();
|
||||
assistantPrefill = lastMsg.content || '';
|
||||
}
|
||||
|
||||
const top64 = b64UrlEncode(JSON.stringify(topMessages));
|
||||
const uniqueId = generateUniqueId('l0');
|
||||
|
||||
const args = {
|
||||
@@ -64,8 +71,12 @@ export async function callLLM(messages, options = {}) {
|
||||
model: DEFAULT_L0_MODEL,
|
||||
};
|
||||
|
||||
// ★ 用 bottomassistant 参数传递 prefill
|
||||
if (assistantPrefill) {
|
||||
args.bottomassistant = assistantPrefill;
|
||||
}
|
||||
|
||||
try {
|
||||
// 非流式直接返回结果
|
||||
const result = await mod.xbgenrawCommand(args, '');
|
||||
return String(result ?? '');
|
||||
} catch (e) {
|
||||
|
||||
@@ -1,52 +1,228 @@
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// query-expansion.js - 完整输入，不截断
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// ============================================================================
|
||||
// query-expansion.js - 检索查询生成器（三层 themes 版）
|
||||
// ============================================================================
|
||||
|
||||
import { callLLM, parseJson } from './llm-service.js';
|
||||
import { xbLog } from '../../../../core/debug-core.js';
|
||||
import { filterText } from '../utils/text-filter.js';
|
||||
import { getContext } from '../../../../../../../extensions.js';
|
||||
import { getSummaryStore } from '../../data/store.js';
|
||||
|
||||
const MODULE_ID = 'query-expansion';
|
||||
const SESSION_ID = 'xb6';
|
||||
|
||||
const SYSTEM_PROMPT = `你是检索词生成器。根据最近对话，输出用于检索历史剧情的关键词。
|
||||
// ============================================================================
|
||||
// 系统提示词
|
||||
// ============================================================================
|
||||
|
||||
只输出JSON：
|
||||
{"e":["显式人物/地名"],"i":["隐含人物/情绪/话题"],"q":["检索短句"]}
|
||||
const SYSTEM_PROMPT = `你是检索查询生成器。根据当前对话上下文，生成用于检索历史剧情的查询语句。
|
||||
|
||||
规则：
|
||||
- e: 对话中明确提到的人名/地名，1-4个
|
||||
- i: 推断出的相关人物/情绪/话题，1-5个
|
||||
- q: 用于向量检索的短句，2-3个，每个15字内
|
||||
- 关注：正在讨论什么、涉及谁、情绪氛围`;
|
||||
|
||||
/**
|
||||
* Query Expansion
|
||||
* @param {Array} messages - 完整消息数组（最后2-3轮）
|
||||
*/
|
||||
export async function expandQuery(messages, options = {}) {
|
||||
const { timeout = 6000 } = options;
|
||||
|
||||
if (!messages?.length) {
|
||||
return { entities: [], implicit: [], queries: [] };
|
||||
## 输出格式（严格JSON）
|
||||
{
|
||||
"focus": ["焦点人物"],
|
||||
"fn": ["叙事功能"],
|
||||
"pt": ["互动模式"],
|
||||
"kw": ["关键词"],
|
||||
"queries": ["DSL查询语句"]
|
||||
}
|
||||
|
||||
// 完整格式化，不截断
|
||||
const input = messages.map(m => {
|
||||
const speaker = m.is_user ? '用户' : (m.name || '角色');
|
||||
const text = filterText(m.mes || '').trim();
|
||||
return `【${speaker}】\n${text}`;
|
||||
}).join('\n\n');
|
||||
## fn（叙事功能）枚举
|
||||
establish=建立设定 | escalate=升级加剧 | reveal=揭示发现 | challenge=挑战试探
|
||||
commit=承诺锁定 | conflict=冲突对抗 | resolve=解决收束 | transform=转变逆转
|
||||
bond=连接羁绊 | break=断裂破坏
|
||||
|
||||
## pt（互动模式）枚举
|
||||
power_down=上对下 | power_up=下对上 | power_equal=对等 | power_contest=争夺
|
||||
asymmetric=信息不对称 | witnessed=有观众 | secluded=隔绝私密
|
||||
ritual=仪式正式 | routine=日常惯例 | triangular=三方介入
|
||||
|
||||
## DSL 查询格式
|
||||
- <act> 主体 -> 动作 (-> 客体)? (在地点)?
|
||||
- <emo> 主体 -> 情绪 (对客体)?
|
||||
- <dec> 主体 -> 决定/承诺 (对客体)?
|
||||
- <rev> 揭示: 内容 (关于客体)?
|
||||
- <ten> 主体A <-> 主体B: 冲突内容
|
||||
- <loc> 场景: 地点/状态
|
||||
|
||||
## 规则
|
||||
- focus: 核心人物，1-4个
|
||||
- fn: 当前对话涉及的叙事功能，1-3个
|
||||
- pt: 当前对话涉及的互动模式，1-3个
|
||||
- kw: 具体关键词，1-4个
|
||||
- queries: 2-4条 DSL 查询
|
||||
|
||||
## 示例
|
||||
|
||||
输入：艾拉说"那把剑...我记得它的重量，在火山口的时候"
|
||||
输出：
|
||||
{
|
||||
"focus": ["艾拉", "古龙"],
|
||||
"fn": ["commit", "bond"],
|
||||
"pt": ["power_down", "ritual"],
|
||||
"kw": ["圣剑", "战斗", "火山口"],
|
||||
"queries": [
|
||||
"<act> 艾拉 -> 战斗/使用圣剑 -> 古龙 [commit/power_down]",
|
||||
"<loc> 场景: 火山口 [ritual]",
|
||||
"<emo> 艾拉 -> 牺牲/决绝 [commit]"
|
||||
]
|
||||
}`;
|
||||
|
||||
// ============================================================================
|
||||
// 上下文构建
|
||||
// ============================================================================
|
||||
|
||||
function getCharacterContext() {
|
||||
const context = getContext();
|
||||
const char = context.characters?.[context.characterId];
|
||||
|
||||
if (!char) {
|
||||
return { name: '', description: '', personality: '' };
|
||||
}
|
||||
|
||||
return {
|
||||
name: char.name || '',
|
||||
description: (char.description || '').slice(0, 500),
|
||||
personality: (char.personality || '').slice(0, 300),
|
||||
};
|
||||
}
|
||||
|
||||
function getPersonaContext() {
|
||||
const context = getContext();
|
||||
|
||||
if (typeof window !== 'undefined' && window.power_user?.persona_description) {
|
||||
return String(window.power_user.persona_description).slice(0, 500);
|
||||
}
|
||||
|
||||
if (context.persona_description) {
|
||||
return String(context.persona_description).slice(0, 500);
|
||||
}
|
||||
|
||||
return '';
|
||||
}
|
||||
|
||||
function getRecentEvents(count = 8) {
|
||||
const store = getSummaryStore();
|
||||
const events = store?.json?.events || [];
|
||||
|
||||
return events
|
||||
.slice(-count)
|
||||
.map(e => {
|
||||
const time = e.timeLabel || '';
|
||||
const title = e.title || '';
|
||||
const participants = (e.participants || []).join('/');
|
||||
const summary = (e.summary || '').replace(/\s*\(#\d+(?:-\d+)?\)\s*$/, '').slice(0, 80);
|
||||
|
||||
return time
|
||||
? `[${time}] ${title || participants}: ${summary}`
|
||||
: `${title || participants}: ${summary}`;
|
||||
});
|
||||
}
|
||||
|
||||
function getRelevantArcs(focusHint = []) {
|
||||
const store = getSummaryStore();
|
||||
const arcs = store?.json?.arcs || [];
|
||||
|
||||
if (!arcs.length) return [];
|
||||
|
||||
const hintSet = new Set(focusHint.map(s => String(s).toLowerCase()));
|
||||
|
||||
const sorted = [...arcs].sort((a, b) => {
|
||||
const aHit = hintSet.has(String(a.name || '').toLowerCase()) ? 1 : 0;
|
||||
const bHit = hintSet.has(String(b.name || '').toLowerCase()) ? 1 : 0;
|
||||
return bHit - aHit;
|
||||
});
|
||||
|
||||
return sorted.slice(0, 4).map(a => {
|
||||
const progress = Math.round((a.progress || 0) * 100);
|
||||
return `${a.name}: ${a.trajectory || '未知状态'} (${progress}%)`;
|
||||
});
|
||||
}
|
||||
|
||||
function extractNamesFromMessages(messages) {
|
||||
const names = new Set();
|
||||
|
||||
for (const m of messages) {
|
||||
if (m.name) names.add(m.name);
|
||||
}
|
||||
|
||||
const text = messages.map(m => m.mes || '').join(' ');
|
||||
const namePattern = /[\u4e00-\u9fff]{2,4}/g;
|
||||
const matches = text.match(namePattern) || [];
|
||||
|
||||
const freq = {};
|
||||
for (const name of matches) {
|
||||
freq[name] = (freq[name] || 0) + 1;
|
||||
}
|
||||
|
||||
Object.entries(freq)
|
||||
.filter(([, count]) => count >= 2)
|
||||
.forEach(([name]) => names.add(name));
|
||||
|
||||
return Array.from(names).slice(0, 6);
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// 主函数
|
||||
// ============================================================================
|
||||
|
||||
export async function expandQuery(messages, options = {}) {
|
||||
const { pendingUserMessage = null, timeout = 6000 } = options;
|
||||
|
||||
if (!messages?.length && !pendingUserMessage) {
|
||||
return { focus: [], fn: [], pt: [], kw: [], queries: [] };
|
||||
}
|
||||
|
||||
const T0 = performance.now();
|
||||
|
||||
const character = getCharacterContext();
|
||||
const persona = getPersonaContext();
|
||||
const nameHints = extractNamesFromMessages(messages || []);
|
||||
const recentEvents = getRecentEvents(8);
|
||||
const arcs = getRelevantArcs(nameHints);
|
||||
|
||||
const dialogueParts = [];
|
||||
|
||||
for (const m of (messages || [])) {
|
||||
const speaker = m.is_user ? '用户' : (m.name || '角色');
|
||||
const text = filterText(m.mes || '').trim();
|
||||
if (text) {
|
||||
dialogueParts.push(`【${speaker}】\n${text.slice(0, 400)}`);
|
||||
}
|
||||
}
|
||||
|
||||
if (pendingUserMessage) {
|
||||
dialogueParts.push(`【用户（刚输入）】\n${filterText(pendingUserMessage).slice(0, 400)}`);
|
||||
}
|
||||
|
||||
const inputParts = [];
|
||||
|
||||
if (character.name) {
|
||||
inputParts.push(`## 当前角色\n${character.name}: ${character.description || character.personality || '无描述'}`);
|
||||
}
|
||||
|
||||
if (persona) {
|
||||
inputParts.push(`## 用户人设\n${persona}`);
|
||||
}
|
||||
|
||||
if (recentEvents.length) {
|
||||
inputParts.push(`## 近期剧情\n${recentEvents.map((e, i) => `${i + 1}. ${e}`).join('\n')}`);
|
||||
}
|
||||
|
||||
if (arcs.length) {
|
||||
inputParts.push(`## 角色状态\n${arcs.join('\n')}`);
|
||||
}
|
||||
|
||||
inputParts.push(`## 最近对话\n${dialogueParts.join('\n\n')}`);
|
||||
|
||||
const input = inputParts.join('\n\n');
|
||||
|
||||
try {
|
||||
const response = await callLLM([
|
||||
{ role: 'system', content: SYSTEM_PROMPT },
|
||||
{ role: 'user', content: input },
|
||||
], {
|
||||
temperature: 0.15,
|
||||
max_tokens: 250,
|
||||
max_tokens: 500,
|
||||
timeout,
|
||||
sessionId: SESSION_ID,
|
||||
});
|
||||
@@ -54,49 +230,104 @@ export async function expandQuery(messages, options = {}) {
|
||||
const parsed = parseJson(response);
|
||||
if (!parsed) {
|
||||
xbLog.warn(MODULE_ID, 'JSON解析失败', response?.slice(0, 200));
|
||||
return { entities: [], implicit: [], queries: [] };
|
||||
return { focus: [], fn: [], pt: [], kw: [], queries: [] };
|
||||
}
|
||||
|
||||
const result = {
|
||||
entities: Array.isArray(parsed.e) ? parsed.e.slice(0, 5) : [],
|
||||
implicit: Array.isArray(parsed.i) ? parsed.i.slice(0, 6) : [],
|
||||
queries: Array.isArray(parsed.q) ? parsed.q.slice(0, 4) : [],
|
||||
focus: Array.isArray(parsed.focus) ? parsed.focus.slice(0, 5) : [],
|
||||
fn: Array.isArray(parsed.fn) ? parsed.fn.slice(0, 4) : [],
|
||||
pt: Array.isArray(parsed.pt) ? parsed.pt.slice(0, 4) : [],
|
||||
kw: Array.isArray(parsed.kw) ? parsed.kw.slice(0, 5) : [],
|
||||
queries: Array.isArray(parsed.queries) ? parsed.queries.slice(0, 5) : [],
|
||||
};
|
||||
|
||||
xbLog.info(MODULE_ID, `完成 (${Math.round(performance.now() - T0)}ms) e=${result.entities.length} i=${result.implicit.length} q=${result.queries.length}`);
|
||||
xbLog.info(MODULE_ID, `完成 (${Math.round(performance.now() - T0)}ms) focus=[${result.focus.join(',')}] fn=[${result.fn.join(',')}]`);
|
||||
return result;
|
||||
|
||||
} catch (e) {
|
||||
xbLog.error(MODULE_ID, '调用失败', e);
|
||||
return { entities: [], implicit: [], queries: [] };
|
||||
return { focus: [], fn: [], pt: [], kw: [], queries: [] };
|
||||
}
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// 缓存
|
||||
// ============================================================================
|
||||
|
||||
const cache = new Map();
|
||||
const CACHE_TTL = 300000;
|
||||
|
||||
function hashMessages(messages) {
|
||||
const text = messages.slice(-2).map(m => (m.mes || '').slice(0, 100)).join('|');
|
||||
function hashMessages(messages, pending = '') {
|
||||
const text = (messages || [])
|
||||
.slice(-3)
|
||||
.map(m => (m.mes || '').slice(0, 100))
|
||||
.join('|') + '|' + (pending || '').slice(0, 100);
|
||||
|
||||
let h = 0;
|
||||
for (let i = 0; i < text.length; i++) h = ((h << 5) - h + text.charCodeAt(i)) | 0;
|
||||
for (let i = 0; i < text.length; i++) {
|
||||
h = ((h << 5) - h + text.charCodeAt(i)) | 0;
|
||||
}
|
||||
return h.toString(36);
|
||||
}
|
||||
|
||||
export async function expandQueryCached(messages, options = {}) {
|
||||
const key = hashMessages(messages);
|
||||
const key = hashMessages(messages, options.pendingUserMessage);
|
||||
const cached = cache.get(key);
|
||||
if (cached && Date.now() - cached.time < CACHE_TTL) return cached.result;
|
||||
|
||||
if (cached && Date.now() - cached.time < CACHE_TTL) {
|
||||
return cached.result;
|
||||
}
|
||||
|
||||
const result = await expandQuery(messages, options);
|
||||
if (result.entities.length || result.queries.length) {
|
||||
if (cache.size > 50) cache.delete(cache.keys().next().value);
|
||||
|
||||
if (result.focus.length || result.queries.length) {
|
||||
if (cache.size > 50) {
|
||||
cache.delete(cache.keys().next().value);
|
||||
}
|
||||
cache.set(key, { result, time: Date.now() });
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// 辅助函数：构建检索文本
|
||||
// ============================================================================
|
||||
|
||||
/**
|
||||
* 将 expansion 结果转换为检索文本
|
||||
* 三层 themes 自然拼入，让向量自动编码
|
||||
*/
|
||||
export function buildSearchText(expansion) {
|
||||
return [...(expansion.entities || []), ...(expansion.implicit || []), ...(expansion.queries || [])]
|
||||
.filter(Boolean).join(' ');
|
||||
const parts = [];
|
||||
|
||||
// focus 人物
|
||||
if (expansion.focus?.length) {
|
||||
parts.push(expansion.focus.join(' '));
|
||||
}
|
||||
|
||||
// fn + pt + kw 合并为标签
|
||||
const tags = [
|
||||
...(expansion.fn || []),
|
||||
...(expansion.pt || []),
|
||||
...(expansion.kw || []),
|
||||
].filter(Boolean);
|
||||
|
||||
if (tags.length) {
|
||||
parts.push(`[${tags.join('/')}]`);
|
||||
}
|
||||
|
||||
// queries
|
||||
if (expansion.queries?.length) {
|
||||
parts.push(...expansion.queries);
|
||||
}
|
||||
|
||||
return parts.filter(Boolean).join(' ').slice(0, 1500);
|
||||
}
|
||||
|
||||
/**
|
||||
* 提取实体列表（兼容旧接口）
|
||||
*/
|
||||
export function getEntitiesFromExpansion(expansion) {
|
||||
return expansion?.focus || [];
|
||||
}
|
||||
|
||||
184
modules/story-summary/vector/llm/reranker.js
Normal file
184
modules/story-summary/vector/llm/reranker.js
Normal file
@@ -0,0 +1,184 @@
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// Reranker - 硅基 bge-reranker-v2-m3
|
||||
// 对候选文档进行精排，过滤与 query 不相关的内容
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
|
||||
import { xbLog } from '../../../../core/debug-core.js';
|
||||
import { getApiKey } from './siliconflow.js';
|
||||
|
||||
const MODULE_ID = 'reranker';
|
||||
const RERANK_URL = 'https://api.siliconflow.cn/v1/rerank';
|
||||
const RERANK_MODEL = 'BAAI/bge-reranker-v2-m3';
|
||||
const DEFAULT_TIMEOUT = 15000;
|
||||
const MAX_DOCUMENTS = 100; // API 限制
|
||||
|
||||
/**
|
||||
* 对文档列表进行 Rerank 精排
|
||||
*
|
||||
* @param {string} query - 查询文本
|
||||
* @param {Array<string>} documents - 文档文本列表
|
||||
* @param {object} options - 选项
|
||||
* @param {number} options.topN - 返回前 N 个结果，默认 40
|
||||
* @param {number} options.timeout - 超时时间，默认 15000ms
|
||||
* @param {AbortSignal} options.signal - 取消信号
|
||||
* @returns {Promise<Array<{index: number, relevance_score: number}>>} 排序后的结果
|
||||
*/
|
||||
export async function rerank(query, documents, options = {}) {
|
||||
const { topN = 40, timeout = DEFAULT_TIMEOUT, signal } = options;
|
||||
|
||||
if (!query?.trim()) {
|
||||
xbLog.warn(MODULE_ID, 'query 为空，跳过 rerank');
|
||||
return documents.map((_, i) => ({ index: i, relevance_score: 0.5 }));
|
||||
}
|
||||
|
||||
if (!documents?.length) {
|
||||
return [];
|
||||
}
|
||||
|
||||
const key = getApiKey();
|
||||
if (!key) {
|
||||
xbLog.warn(MODULE_ID, '未配置 API Key，跳过 rerank');
|
||||
return documents.map((_, i) => ({ index: i, relevance_score: 0.5 }));
|
||||
}
|
||||
|
||||
// 截断超长文档列表
|
||||
const truncatedDocs = documents.slice(0, MAX_DOCUMENTS);
|
||||
if (documents.length > MAX_DOCUMENTS) {
|
||||
xbLog.warn(MODULE_ID, `文档数 ${documents.length} 超过限制 ${MAX_DOCUMENTS}，已截断`);
|
||||
}
|
||||
|
||||
// 过滤空文档，记录原始索引
|
||||
const validDocs = [];
|
||||
const indexMap = []; // validDocs index → original index
|
||||
|
||||
for (let i = 0; i < truncatedDocs.length; i++) {
|
||||
const text = String(truncatedDocs[i] || '').trim();
|
||||
if (text) {
|
||||
validDocs.push(text);
|
||||
indexMap.push(i);
|
||||
}
|
||||
}
|
||||
|
||||
if (!validDocs.length) {
|
||||
xbLog.warn(MODULE_ID, '无有效文档，跳过 rerank');
|
||||
return [];
|
||||
}
|
||||
|
||||
const controller = new AbortController();
|
||||
const timeoutId = setTimeout(() => controller.abort(), timeout);
|
||||
|
||||
try {
|
||||
const T0 = performance.now();
|
||||
|
||||
const response = await fetch(RERANK_URL, {
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Authorization': `Bearer ${key}`,
|
||||
'Content-Type': 'application/json',
|
||||
},
|
||||
body: JSON.stringify({
|
||||
model: RERANK_MODEL,
|
||||
query: query.slice(0, 1000), // 限制 query 长度
|
||||
documents: validDocs,
|
||||
top_n: Math.min(topN, validDocs.length),
|
||||
return_documents: false,
|
||||
}),
|
||||
signal: signal || controller.signal,
|
||||
});
|
||||
|
||||
clearTimeout(timeoutId);
|
||||
|
||||
if (!response.ok) {
|
||||
const errorText = await response.text().catch(() => '');
|
||||
throw new Error(`Rerank API ${response.status}: ${errorText.slice(0, 200)}`);
|
||||
}
|
||||
|
||||
const data = await response.json();
|
||||
const results = data.results || [];
|
||||
|
||||
// 映射回原始索引
|
||||
const mapped = results.map(r => ({
|
||||
index: indexMap[r.index],
|
||||
relevance_score: r.relevance_score ?? 0,
|
||||
}));
|
||||
|
||||
const elapsed = Math.round(performance.now() - T0);
|
||||
xbLog.info(MODULE_ID, `Rerank 完成: ${validDocs.length} docs → ${results.length} selected (${elapsed}ms)`);
|
||||
|
||||
return mapped;
|
||||
|
||||
} catch (e) {
|
||||
clearTimeout(timeoutId);
|
||||
|
||||
if (e?.name === 'AbortError') {
|
||||
xbLog.warn(MODULE_ID, 'Rerank 超时或取消');
|
||||
} else {
|
||||
xbLog.error(MODULE_ID, 'Rerank 失败', e);
|
||||
}
|
||||
|
||||
// 降级：返回原顺序，分数均匀分布
|
||||
return documents.slice(0, topN).map((_, i) => ({
|
||||
index: i,
|
||||
relevance_score: 1 - (i / documents.length) * 0.5,
|
||||
}));
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* 对 chunk 对象列表进行 Rerank
|
||||
*
|
||||
* @param {string} query - 查询文本
|
||||
* @param {Array<object>} chunks - chunk 对象列表，需要有 text 字段
|
||||
* @param {object} options - 选项
|
||||
* @returns {Promise<Array<object>>} 排序后的 chunk 列表，带 _rerankScore 字段
|
||||
*/
|
||||
export async function rerankChunks(query, chunks, options = {}) {
|
||||
const { topN = 40, minScore = 0.1 } = options;
|
||||
|
||||
if (!chunks?.length) return [];
|
||||
if (chunks.length <= topN) {
|
||||
// 数量不超限，仍然 rerank 以获取分数，但不过滤
|
||||
const texts = chunks.map(c => c.text || c.semantic || '');
|
||||
const results = await rerank(query, texts, { topN: chunks.length, ...options });
|
||||
|
||||
const scoreMap = new Map(results.map(r => [r.index, r.relevance_score]));
|
||||
return chunks.map((c, i) => ({
|
||||
...c,
|
||||
_rerankScore: scoreMap.get(i) ?? 0.5,
|
||||
})).sort((a, b) => b._rerankScore - a._rerankScore);
|
||||
}
|
||||
|
||||
const texts = chunks.map(c => c.text || c.semantic || '');
|
||||
const results = await rerank(query, texts, { topN, ...options });
|
||||
|
||||
// 过滤低分 + 排序
|
||||
const selected = results
|
||||
.filter(r => r.relevance_score >= minScore)
|
||||
.sort((a, b) => b.relevance_score - a.relevance_score)
|
||||
.map(r => ({
|
||||
...chunks[r.index],
|
||||
_rerankScore: r.relevance_score,
|
||||
}));
|
||||
|
||||
return selected;
|
||||
}
|
||||
|
||||
/**
|
||||
* 测试 Rerank 服务连接
|
||||
*/
|
||||
export async function testRerankService() {
|
||||
const key = getApiKey();
|
||||
if (!key) {
|
||||
throw new Error('请配置硅基 API Key');
|
||||
}
|
||||
|
||||
try {
|
||||
const results = await rerank('测试查询', ['测试文档1', '测试文档2'], { topN: 2 });
|
||||
return {
|
||||
success: true,
|
||||
message: `连接成功，返回 ${results.length} 个结果`,
|
||||
};
|
||||
} catch (e) {
|
||||
throw new Error(`连接失败: ${e.message}`);
|
||||
}
|
||||
}
|
||||
@@ -1,9 +1,11 @@
|
||||
// ============================================================================
|
||||
// state-integration.js - L0 记忆锚点管理
|
||||
// 支持增量提取、清空、取消
|
||||
// ============================================================================
|
||||
// state-integration.js - L0 状态层集成
|
||||
// Phase 1: 批量 LLM 提取（只存文本）
|
||||
// Phase 2: 统一向量化（提取完成后）
|
||||
// ============================================================================
|
||||
|
||||
import { getContext } from '../../../../../../../extensions.js';
|
||||
import { saveMetadataDebounced } from '../../../../../../../extensions.js';
|
||||
import { xbLog } from '../../../../core/debug-core.js';
|
||||
import {
|
||||
saveStateAtoms,
|
||||
@@ -26,9 +28,15 @@ import { filterText } from '../utils/text-filter.js';
|
||||
|
||||
const MODULE_ID = 'state-integration';
|
||||
|
||||
// ★ 并发配置
|
||||
const CONCURRENCY = 30;
|
||||
const STAGGER_DELAY = 30;
|
||||
|
||||
let initialized = false;
|
||||
let extractionCancelled = false;
|
||||
|
||||
export function cancelL0Extraction() {
|
||||
extractionCancelled = true;
|
||||
cancelBatchExtraction();
|
||||
}
|
||||
|
||||
@@ -53,6 +61,7 @@ export async function getAnchorStats() {
|
||||
return { extracted: 0, total: 0, pending: 0, empty: 0, fail: 0 };
|
||||
}
|
||||
|
||||
// 统计 AI 楼层
|
||||
const aiFloors = [];
|
||||
for (let i = 0; i < chat.length; i++) {
|
||||
if (!chat[i]?.is_user) aiFloors.push(i);
|
||||
@@ -71,14 +80,20 @@ export async function getAnchorStats() {
|
||||
}
|
||||
|
||||
const total = aiFloors.length;
|
||||
const completed = ok + empty;
|
||||
const pending = Math.max(0, total - completed);
|
||||
const processed = ok + empty + fail;
|
||||
const pending = Math.max(0, total - processed);
|
||||
|
||||
return { extracted: completed, total, pending, empty, fail };
|
||||
return {
|
||||
extracted: ok + empty,
|
||||
total,
|
||||
pending,
|
||||
empty,
|
||||
fail
|
||||
};
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// 增量提取
|
||||
// 增量提取 - Phase 1 提取文本，Phase 2 统一向量化
|
||||
// ============================================================================
|
||||
|
||||
function buildL0InputText(userMessage, aiMessage) {
|
||||
@@ -102,6 +117,9 @@ export async function incrementalExtractAtoms(chatId, chat, onProgress) {
|
||||
const vectorCfg = getVectorConfig();
|
||||
if (!vectorCfg?.enabled) return { built: 0 };
|
||||
|
||||
// ★ 重置取消标志
|
||||
extractionCancelled = false;
|
||||
|
||||
const pendingPairs = [];
|
||||
|
||||
for (let i = 0; i < chat.length; i++) {
|
||||
@@ -109,6 +127,7 @@ export async function incrementalExtractAtoms(chatId, chat, onProgress) {
|
||||
if (!msg || msg.is_user) continue;
|
||||
|
||||
const st = getL0FloorStatus(i);
|
||||
// ★ 只跳过 ok 和 empty，fail 的可以重试
|
||||
if (st?.status === 'ok' || st?.status === 'empty') {
|
||||
continue;
|
||||
}
|
||||
@@ -125,23 +144,47 @@ export async function incrementalExtractAtoms(chatId, chat, onProgress) {
|
||||
}
|
||||
|
||||
if (!pendingPairs.length) {
|
||||
onProgress?.(0, 0, '已全部提取');
|
||||
onProgress?.('已全部提取', 0, 0);
|
||||
return { built: 0 };
|
||||
}
|
||||
|
||||
xbLog.info(MODULE_ID, `增量 L0 提取：pending=${pendingPairs.length}`);
|
||||
xbLog.info(MODULE_ID, `增量 L0 提取：pending=${pendingPairs.length}, concurrency=${CONCURRENCY}`);
|
||||
|
||||
let completed = 0;
|
||||
let failed = 0;
|
||||
const total = pendingPairs.length;
|
||||
let builtAtoms = 0;
|
||||
|
||||
for (const pair of pendingPairs) {
|
||||
// ★ Phase 1: 收集所有新提取的 atoms（不向量化）
|
||||
const allNewAtoms = [];
|
||||
|
||||
// ★ 30 并发批次处理
|
||||
for (let i = 0; i < pendingPairs.length; i += CONCURRENCY) {
|
||||
// ★ 检查取消
|
||||
if (extractionCancelled) {
|
||||
xbLog.info(MODULE_ID, `用户取消，已完成 ${completed}/${total}`);
|
||||
break;
|
||||
}
|
||||
|
||||
const batch = pendingPairs.slice(i, i + CONCURRENCY);
|
||||
|
||||
const promises = batch.map((pair, idx) => (async () => {
|
||||
// 首批错开启动，避免瞬间打满
|
||||
if (i === 0) {
|
||||
await new Promise(r => setTimeout(r, idx * STAGGER_DELAY));
|
||||
}
|
||||
|
||||
// 再次检查取消
|
||||
if (extractionCancelled) return;
|
||||
|
||||
const floor = pair.aiFloor;
|
||||
const prev = getL0FloorStatus(floor);
|
||||
|
||||
try {
|
||||
const atoms = await extractAtomsForRound(pair.userMsg, pair.aiMsg, floor, { timeout: 20000 });
|
||||
|
||||
if (extractionCancelled) return;
|
||||
|
||||
if (atoms == null) {
|
||||
throw new Error('llm_failed');
|
||||
}
|
||||
@@ -151,28 +194,59 @@ export async function incrementalExtractAtoms(chatId, chat, onProgress) {
|
||||
} else {
|
||||
atoms.forEach(a => a.chatId = chatId);
|
||||
saveStateAtoms(atoms);
|
||||
await vectorizeAtoms(chatId, atoms);
|
||||
// ★ Phase 1: 只收集，不向量化
|
||||
allNewAtoms.push(...atoms);
|
||||
|
||||
setL0FloorStatus(floor, { status: 'ok', atoms: atoms.length });
|
||||
builtAtoms += atoms.length;
|
||||
}
|
||||
} catch (e) {
|
||||
if (extractionCancelled) return;
|
||||
|
||||
setL0FloorStatus(floor, {
|
||||
status: 'fail',
|
||||
attempts: (prev?.attempts || 0) + 1,
|
||||
reason: String(e?.message || e).replace(/\s+/g, ' ').slice(0, 120),
|
||||
});
|
||||
failed++;
|
||||
} finally {
|
||||
if (!extractionCancelled) {
|
||||
completed++;
|
||||
onProgress?.(`L0: ${completed}/${total}`, completed, total);
|
||||
onProgress?.(`提取: ${completed}/${total}`, completed, total);
|
||||
}
|
||||
}
|
||||
})());
|
||||
|
||||
await Promise.all(promises);
|
||||
|
||||
// 批次间短暂间隔
|
||||
if (i + CONCURRENCY < pendingPairs.length && !extractionCancelled) {
|
||||
await new Promise(r => setTimeout(r, 30));
|
||||
}
|
||||
}
|
||||
|
||||
xbLog.info(MODULE_ID, `增量 L0 完成：atoms=${builtAtoms}, floors=${pendingPairs.length}`);
|
||||
// ★ 立即保存文本，不要等防抖
|
||||
try {
|
||||
saveMetadataDebounced?.();
|
||||
} catch { }
|
||||
|
||||
// ★ Phase 2: 统一向量化所有新提取的 atoms
|
||||
if (allNewAtoms.length > 0 && !extractionCancelled) {
|
||||
onProgress?.(`向量化 L0: 0/${allNewAtoms.length}`, 0, allNewAtoms.length);
|
||||
await vectorizeAtoms(chatId, allNewAtoms, (current, total) => {
|
||||
onProgress?.(`向量化 L0: ${current}/${total}`, current, total);
|
||||
});
|
||||
}
|
||||
|
||||
xbLog.info(MODULE_ID, `L0 ${extractionCancelled ? '已取消' : '完成'}：atoms=${builtAtoms}, completed=${completed}/${total}, failed=${failed}`);
|
||||
return { built: builtAtoms };
|
||||
}
|
||||
|
||||
async function vectorizeAtoms(chatId, atoms) {
|
||||
// ============================================================================
|
||||
// 向量化（支持进度回调）
|
||||
// ============================================================================
|
||||
|
||||
async function vectorizeAtoms(chatId, atoms, onProgress) {
|
||||
if (!atoms?.length) return;
|
||||
|
||||
const vectorCfg = getVectorConfig();
|
||||
@@ -180,14 +254,27 @@ async function vectorizeAtoms(chatId, atoms) {
|
||||
|
||||
const texts = atoms.map(a => a.semantic);
|
||||
const fingerprint = getEngineFingerprint(vectorCfg);
|
||||
const batchSize = 20;
|
||||
|
||||
try {
|
||||
const vectors = await embed(texts, { timeout: 30000 });
|
||||
const allVectors = [];
|
||||
|
||||
const items = atoms.map((a, i) => ({
|
||||
for (let i = 0; i < texts.length; i += batchSize) {
|
||||
if (extractionCancelled) break;
|
||||
|
||||
const batch = texts.slice(i, i + batchSize);
|
||||
const vectors = await embed(batch, { timeout: 30000 });
|
||||
allVectors.push(...vectors);
|
||||
|
||||
onProgress?.(allVectors.length, texts.length);
|
||||
}
|
||||
|
||||
if (extractionCancelled) return;
|
||||
|
||||
const items = atoms.slice(0, allVectors.length).map((a, i) => ({
|
||||
atomId: a.atomId,
|
||||
floor: a.floor,
|
||||
vector: vectors[i],
|
||||
vector: allVectors[i],
|
||||
}));
|
||||
|
||||
await saveStateVectors(chatId, items, fingerprint);
|
||||
@@ -207,11 +294,17 @@ export async function clearAllAtomsAndVectors(chatId) {
|
||||
if (chatId) {
|
||||
await clearStateVectors(chatId);
|
||||
}
|
||||
|
||||
// ★ 立即保存
|
||||
try {
|
||||
saveMetadataDebounced?.();
|
||||
} catch { }
|
||||
|
||||
xbLog.info(MODULE_ID, '已清空所有记忆锚点');
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// 实时增量（AI 消息后触发）- 保留原有逻辑
|
||||
// 实时增量（AI 消息后触发）- 保持不变
|
||||
// ============================================================================
|
||||
|
||||
let extractionQueue = [];
|
||||
@@ -245,7 +338,9 @@ async function processQueue() {
|
||||
|
||||
atoms.forEach(a => a.chatId = chatId);
|
||||
saveStateAtoms(atoms);
|
||||
await vectorizeAtoms(chatId, atoms);
|
||||
|
||||
// 单楼实时处理：立即向量化
|
||||
await vectorizeAtomsSimple(chatId, atoms);
|
||||
|
||||
xbLog.info(MODULE_ID, `floor ${aiFloor}: ${atoms.length} atoms 已存储`);
|
||||
} catch (e) {
|
||||
@@ -256,6 +351,31 @@ async function processQueue() {
|
||||
isProcessing = false;
|
||||
}
|
||||
|
||||
// 简单向量化（无进度回调，用于单楼实时处理）
|
||||
async function vectorizeAtomsSimple(chatId, atoms) {
|
||||
if (!atoms?.length) return;
|
||||
|
||||
const vectorCfg = getVectorConfig();
|
||||
if (!vectorCfg?.enabled) return;
|
||||
|
||||
const texts = atoms.map(a => a.semantic);
|
||||
const fingerprint = getEngineFingerprint(vectorCfg);
|
||||
|
||||
try {
|
||||
const vectors = await embed(texts, { timeout: 30000 });
|
||||
|
||||
const items = atoms.map((a, i) => ({
|
||||
atomId: a.atomId,
|
||||
floor: a.floor,
|
||||
vector: vectors[i],
|
||||
}));
|
||||
|
||||
await saveStateVectors(chatId, items, fingerprint);
|
||||
} catch (e) {
|
||||
xbLog.error(MODULE_ID, 'L0 向量化失败', e);
|
||||
}
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// 回滚钩子
|
||||
// ============================================================================
|
||||
@@ -301,7 +421,7 @@ export async function rebuildStateVectors(chatId, vectorCfg) {
|
||||
xbLog.info(MODULE_ID, `重建 L0 向量: ${atoms.length} 条 atom`);
|
||||
|
||||
await clearStateVectors(chatId);
|
||||
await vectorizeAtoms(chatId, atoms);
|
||||
await vectorizeAtomsSimple(chatId, atoms);
|
||||
|
||||
return { built: atoms.length };
|
||||
}
|
||||
|
||||
@@ -131,16 +131,44 @@ export function stateToVirtualChunks(l0Results) {
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
|
||||
/**
|
||||
* 合并 L0 和 L1 chunks，每楼层最多保留 limit 条
|
||||
* @param {Array} l0Chunks - 虚拟 chunks（已按相似度排序）
|
||||
* @param {Array} l1Chunks - 真实 chunks（已按相似度排序）
|
||||
* 合并 L0 和 L1 chunks
|
||||
* @param {Array} l0Chunks - L0 虚拟 chunks（带 similarity）
|
||||
* @param {Array} l1Chunks - L1 真实 chunks（无 similarity）
|
||||
* @param {number} limit - 每楼层上限
|
||||
* @returns {Array} 合并后的 chunks
|
||||
*/
|
||||
export function mergeAndSparsify(l0Chunks, l1Chunks, limit = 2) {
|
||||
// 构建 L0 楼层 → 最高 similarity 映射
|
||||
const floorSimilarity = new Map();
|
||||
for (const c of (l0Chunks || [])) {
|
||||
const existing = floorSimilarity.get(c.floor) || 0;
|
||||
if ((c.similarity || 0) > existing) {
|
||||
floorSimilarity.set(c.floor, c.similarity || 0);
|
||||
}
|
||||
}
|
||||
|
||||
// L1 继承所属楼层的 L0 similarity
|
||||
const l1WithScore = (l1Chunks || []).map(c => ({
|
||||
...c,
|
||||
similarity: floorSimilarity.get(c.floor) || 0.5,
|
||||
}));
|
||||
|
||||
// 合并并按相似度排序
|
||||
const all = [...(l0Chunks || []), ...(l1Chunks || [])]
|
||||
.sort((a, b) => b.similarity - a.similarity);
|
||||
const all = [...(l0Chunks || []), ...l1WithScore]
|
||||
.sort((a, b) => {
|
||||
// 相似度优先
|
||||
const simDiff = (b.similarity || 0) - (a.similarity || 0);
|
||||
if (Math.abs(simDiff) > 0.01) return simDiff;
|
||||
|
||||
// 同楼层：L0 优先于 L1
|
||||
if (a.floor === b.floor) {
|
||||
if (a.isL0 && !b.isL0) return -1;
|
||||
if (!a.isL0 && b.isL0) return 1;
|
||||
}
|
||||
|
||||
// 按楼层升序
|
||||
return a.floor - b.floor;
|
||||
});
|
||||
|
||||
// 每楼层稀疏去重
|
||||
const byFloor = new Map();
|
||||
@@ -153,8 +181,9 @@ export function mergeAndSparsify(l0Chunks, l1Chunks, limit = 2) {
|
||||
}
|
||||
}
|
||||
|
||||
// 扁平化并保持相似度排序
|
||||
// 扁平化并保持排序
|
||||
return Array.from(byFloor.values())
|
||||
.flat()
|
||||
.sort((a, b) => b.similarity - a.similarity);
|
||||
.sort((a, b) => (b.similarity || 0) - (a.similarity || 0));
|
||||
}
|
||||
|
||||
|
||||
388
modules/story-summary/vector/retrieval/metrics.js
Normal file
388
modules/story-summary/vector/retrieval/metrics.js
Normal file
@@ -0,0 +1,388 @@
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// Story Summary - Metrics Collector
|
||||
// 召回质量指标收集与格式化
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
|
||||
/**
|
||||
* 创建空的指标对象
|
||||
*/
|
||||
export function createMetrics() {
|
||||
return {
|
||||
// L0 Query Understanding
|
||||
l0: {
|
||||
needRecall: false,
|
||||
intent: '',
|
||||
focusEntities: [],
|
||||
queries: [],
|
||||
implicitTopics: [],
|
||||
queryExpansionTime: 0,
|
||||
atomsMatched: 0,
|
||||
floorsHit: 0,
|
||||
topAtoms: [],
|
||||
},
|
||||
|
||||
// L1 Constraints (Facts)
|
||||
l1: {
|
||||
factsTotal: 0,
|
||||
factsInjected: 0,
|
||||
factsFiltered: 0,
|
||||
tokens: 0,
|
||||
samples: [],
|
||||
},
|
||||
|
||||
// L2 Narrative Retrieval
|
||||
l2: {
|
||||
eventsInStore: 0,
|
||||
eventsConsidered: 0,
|
||||
eventsSelected: 0,
|
||||
byRecallType: { direct: 0, causal: 0, context: 0 },
|
||||
similarityDistribution: { min: 0, max: 0, mean: 0, median: 0 },
|
||||
entityFilterStats: null,
|
||||
causalChainDepth: 0,
|
||||
causalEventsCount: 0,
|
||||
entitiesLoaded: 0,
|
||||
entityNames: [],
|
||||
retrievalTime: 0,
|
||||
},
|
||||
|
||||
// L3 Evidence Assembly
|
||||
l3: {
|
||||
floorsFromL0: 0,
|
||||
// 候选规模（rerank 前）
|
||||
chunksInRange: 0,
|
||||
chunksInRangeByType: { l0Virtual: 0, l1Real: 0 },
|
||||
// 最终注入（rerank + sparse 后）
|
||||
chunksSelected: 0,
|
||||
chunksSelectedByType: { l0Virtual: 0, l1Real: 0 },
|
||||
// 上下文配对
|
||||
contextPairsAdded: 0,
|
||||
tokens: 0,
|
||||
assemblyTime: 0,
|
||||
// Rerank 相关
|
||||
rerankApplied: false,
|
||||
beforeRerank: 0,
|
||||
afterRerank: 0,
|
||||
rerankTime: 0,
|
||||
rerankScoreDistribution: null,
|
||||
},
|
||||
|
||||
// L4 Formatting
|
||||
l4: {
|
||||
sectionsIncluded: [],
|
||||
formattingTime: 0,
|
||||
},
|
||||
|
||||
// Budget Summary
|
||||
budget: {
|
||||
total: 0,
|
||||
limit: 0,
|
||||
utilization: 0,
|
||||
breakdown: {
|
||||
constraints: 0,
|
||||
events: 0,
|
||||
entities: 0,
|
||||
chunks: 0,
|
||||
recentOrphans: 0,
|
||||
arcs: 0,
|
||||
},
|
||||
},
|
||||
|
||||
// Total Timing
|
||||
timing: {
|
||||
queryExpansion: 0,
|
||||
l0Search: 0,
|
||||
l1Constraints: 0,
|
||||
l2Retrieval: 0,
|
||||
l3Retrieval: 0,
|
||||
l3Rerank: 0,
|
||||
l3Assembly: 0,
|
||||
l4Formatting: 0,
|
||||
total: 0,
|
||||
},
|
||||
|
||||
// Quality Indicators
|
||||
quality: {
|
||||
constraintCoverage: 100,
|
||||
eventPrecisionProxy: 0,
|
||||
evidenceDensity: 0,
|
||||
potentialIssues: [],
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 计算相似度分布统计
|
||||
*/
|
||||
export function calcSimilarityStats(similarities) {
|
||||
if (!similarities?.length) {
|
||||
return { min: 0, max: 0, mean: 0, median: 0 };
|
||||
}
|
||||
|
||||
const sorted = [...similarities].sort((a, b) => a - b);
|
||||
const sum = sorted.reduce((a, b) => a + b, 0);
|
||||
|
||||
return {
|
||||
min: Number(sorted[0].toFixed(3)),
|
||||
max: Number(sorted[sorted.length - 1].toFixed(3)),
|
||||
mean: Number((sum / sorted.length).toFixed(3)),
|
||||
median: Number(sorted[Math.floor(sorted.length / 2)].toFixed(3)),
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式化指标为可读日志
|
||||
*/
|
||||
export function formatMetricsLog(metrics) {
|
||||
const m = metrics;
|
||||
const lines = [];
|
||||
|
||||
lines.push('');
|
||||
lines.push('═══════════════════════════════════════════════════════════════════');
|
||||
lines.push(' Recall Metrics Report ');
|
||||
lines.push('═══════════════════════════════════════════════════════════════════');
|
||||
lines.push('');
|
||||
|
||||
// L0 Query Understanding
|
||||
lines.push('[L0] Query Understanding');
|
||||
lines.push(`├─ need_recall: ${m.l0.needRecall}`);
|
||||
if (m.l0.needRecall) {
|
||||
lines.push(`├─ intent: ${m.l0.intent || 'mixed'}`);
|
||||
lines.push(`├─ focus_entities: [${(m.l0.focusEntities || []).join(', ')}]`);
|
||||
lines.push(`├─ queries: [${(m.l0.queries || []).slice(0, 3).join(', ')}]`);
|
||||
lines.push(`├─ query_expansion_time: ${m.l0.queryExpansionTime}ms`);
|
||||
lines.push(`├─ atoms_matched: ${m.l0.atomsMatched || 0}`);
|
||||
lines.push(`└─ floors_hit: ${m.l0.floorsHit || 0}`);
|
||||
}
|
||||
lines.push('');
|
||||
|
||||
// L1 Constraints
|
||||
lines.push('[L1] Constraints (Facts)');
|
||||
lines.push(`├─ facts_total: ${m.l1.factsTotal}`);
|
||||
lines.push(`├─ facts_filtered: ${m.l1.factsFiltered || 0}`);
|
||||
lines.push(`├─ facts_injected: ${m.l1.factsInjected}`);
|
||||
lines.push(`├─ tokens: ${m.l1.tokens}`);
|
||||
if (m.l1.samples && m.l1.samples.length > 0) {
|
||||
lines.push(`└─ samples: "${m.l1.samples.slice(0, 2).join('", "')}"`);
|
||||
}
|
||||
lines.push('');
|
||||
|
||||
// L2 Narrative Retrieval
|
||||
lines.push('[L2] Narrative Retrieval');
|
||||
lines.push(`├─ events_in_store: ${m.l2.eventsInStore}`);
|
||||
lines.push(`├─ events_considered: ${m.l2.eventsConsidered}`);
|
||||
|
||||
if (m.l2.entityFilterStats) {
|
||||
const ef = m.l2.entityFilterStats;
|
||||
lines.push(`├─ entity_filter:`);
|
||||
lines.push(`│ ├─ focus_entities: [${(ef.focusEntities || []).join(', ')}]`);
|
||||
lines.push(`│ ├─ before_filter: ${ef.before}`);
|
||||
lines.push(`│ ├─ after_filter: ${ef.after}`);
|
||||
lines.push(`│ └─ filtered_out: ${ef.filtered}`);
|
||||
}
|
||||
|
||||
lines.push(`├─ events_selected: ${m.l2.eventsSelected}`);
|
||||
lines.push(`├─ by_recall_type:`);
|
||||
lines.push(`│ ├─ direct: ${m.l2.byRecallType.direct}`);
|
||||
lines.push(`│ ├─ causal: ${m.l2.byRecallType.causal}`);
|
||||
lines.push(`│ └─ context: ${m.l2.byRecallType.context}`);
|
||||
|
||||
const sim = m.l2.similarityDistribution;
|
||||
if (sim && sim.max > 0) {
|
||||
lines.push(`├─ similarity_distribution:`);
|
||||
lines.push(`│ ├─ min: ${sim.min}`);
|
||||
lines.push(`│ ├─ max: ${sim.max}`);
|
||||
lines.push(`│ ├─ mean: ${sim.mean}`);
|
||||
lines.push(`│ └─ median: ${sim.median}`);
|
||||
}
|
||||
|
||||
lines.push(`├─ causal_chain: depth=${m.l2.causalChainDepth}, events=${m.l2.causalEventsCount}`);
|
||||
lines.push(`├─ entities_loaded: ${m.l2.entitiesLoaded} [${(m.l2.entityNames || []).join(', ')}]`);
|
||||
lines.push(`└─ retrieval_time: ${m.l2.retrievalTime}ms`);
|
||||
lines.push('');
|
||||
|
||||
// L3 Evidence Assembly
|
||||
lines.push('[L3] Evidence Assembly');
|
||||
lines.push(`├─ floors_from_l0: ${m.l3.floorsFromL0}`);
|
||||
|
||||
// 候选规模
|
||||
lines.push(`├─ chunks_in_range: ${m.l3.chunksInRange}`);
|
||||
if (m.l3.chunksInRangeByType) {
|
||||
const cir = m.l3.chunksInRangeByType;
|
||||
lines.push(`│ ├─ l0_virtual: ${cir.l0Virtual || 0}`);
|
||||
lines.push(`│ └─ l1_real: ${cir.l1Real || 0}`);
|
||||
}
|
||||
|
||||
// Rerank 信息
|
||||
if (m.l3.rerankApplied) {
|
||||
lines.push(`├─ rerank_applied: true`);
|
||||
lines.push(`│ ├─ before: ${m.l3.beforeRerank}`);
|
||||
lines.push(`│ ├─ after: ${m.l3.afterRerank}`);
|
||||
lines.push(`│ └─ time: ${m.l3.rerankTime}ms`);
|
||||
if (m.l3.rerankScoreDistribution) {
|
||||
const rd = m.l3.rerankScoreDistribution;
|
||||
lines.push(`├─ rerank_scores: min=${rd.min}, max=${rd.max}, mean=${rd.mean}`);
|
||||
}
|
||||
} else {
|
||||
lines.push(`├─ rerank_applied: false`);
|
||||
}
|
||||
|
||||
// 最终注入规模
|
||||
lines.push(`├─ chunks_selected: ${m.l3.chunksSelected}`);
|
||||
if (m.l3.chunksSelectedByType) {
|
||||
const cs = m.l3.chunksSelectedByType;
|
||||
lines.push(`│ ├─ l0_virtual: ${cs.l0Virtual || 0}`);
|
||||
lines.push(`│ └─ l1_real: ${cs.l1Real || 0}`);
|
||||
}
|
||||
|
||||
lines.push(`├─ context_pairs_added: ${m.l3.contextPairsAdded}`);
|
||||
lines.push(`├─ tokens: ${m.l3.tokens}`);
|
||||
lines.push(`└─ assembly_time: ${m.l3.assemblyTime}ms`);
|
||||
lines.push('');
|
||||
|
||||
// L4 Formatting
|
||||
lines.push('[L4] Prompt Formatting');
|
||||
lines.push(`├─ sections: [${(m.l4.sectionsIncluded || []).join(', ')}]`);
|
||||
lines.push(`└─ formatting_time: ${m.l4.formattingTime}ms`);
|
||||
lines.push('');
|
||||
|
||||
// Budget Summary
|
||||
lines.push('[Budget Summary]');
|
||||
lines.push(`├─ total_tokens: ${m.budget.total}`);
|
||||
lines.push(`├─ budget_limit: ${m.budget.limit}`);
|
||||
lines.push(`├─ utilization: ${m.budget.utilization}%`);
|
||||
lines.push(`└─ breakdown:`);
|
||||
const bd = m.budget.breakdown || {};
|
||||
lines.push(` ├─ constraints (L1): ${bd.constraints || 0}`);
|
||||
lines.push(` ├─ events (L2): ${bd.events || 0}`);
|
||||
lines.push(` ├─ chunks (L3): ${bd.chunks || 0}`);
|
||||
lines.push(` ├─ recent_orphans: ${bd.recentOrphans || 0}`);
|
||||
lines.push(` └─ arcs: ${bd.arcs || 0}`);
|
||||
lines.push('');
|
||||
|
||||
// Timing
|
||||
lines.push('[Timing]');
|
||||
lines.push(`├─ query_expansion: ${m.timing.queryExpansion}ms`);
|
||||
lines.push(`├─ l0_search: ${m.timing.l0Search}ms`);
|
||||
lines.push(`├─ l1_constraints: ${m.timing.l1Constraints}ms`);
|
||||
lines.push(`├─ l2_retrieval: ${m.timing.l2Retrieval}ms`);
|
||||
lines.push(`├─ l3_retrieval: ${m.timing.l3Retrieval}ms`);
|
||||
if (m.timing.l3Rerank > 0) {
|
||||
lines.push(`├─ l3_rerank: ${m.timing.l3Rerank}ms`);
|
||||
}
|
||||
lines.push(`├─ l3_assembly: ${m.timing.l3Assembly}ms`);
|
||||
lines.push(`├─ l4_formatting: ${m.timing.l4Formatting}ms`);
|
||||
lines.push(`└─ total: ${m.timing.total}ms`);
|
||||
lines.push('');
|
||||
|
||||
// Quality Indicators
|
||||
lines.push('[Quality Indicators]');
|
||||
lines.push(`├─ constraint_coverage: ${m.quality.constraintCoverage}%`);
|
||||
lines.push(`├─ event_precision_proxy: ${m.quality.eventPrecisionProxy}`);
|
||||
lines.push(`├─ evidence_density: ${m.quality.evidenceDensity}%`);
|
||||
|
||||
if (m.quality.potentialIssues && m.quality.potentialIssues.length > 0) {
|
||||
lines.push(`└─ potential_issues:`);
|
||||
m.quality.potentialIssues.forEach((issue, i) => {
|
||||
const prefix = i === m.quality.potentialIssues.length - 1 ? ' └─' : ' ├─';
|
||||
lines.push(`${prefix} ⚠ ${issue}`);
|
||||
});
|
||||
} else {
|
||||
lines.push(`└─ potential_issues: none`);
|
||||
}
|
||||
|
||||
lines.push('');
|
||||
lines.push('═══════════════════════════════════════════════════════════════════');
|
||||
lines.push('');
|
||||
|
||||
return lines.join('\n');
|
||||
}
|
||||
|
||||
/**
|
||||
* 检测潜在问题
|
||||
*/
|
||||
export function detectIssues(metrics) {
|
||||
const issues = [];
|
||||
const m = metrics;
|
||||
|
||||
// 召回比例问题
|
||||
if (m.l2.eventsConsidered > 0) {
|
||||
const selectRatio = m.l2.eventsSelected / m.l2.eventsConsidered;
|
||||
if (selectRatio < 0.1) {
|
||||
issues.push(`Event selection ratio too low (${(selectRatio * 100).toFixed(1)}%) - threshold may be too high`);
|
||||
}
|
||||
if (selectRatio > 0.6 && m.l2.eventsConsidered > 10) {
|
||||
issues.push(`Event selection ratio high (${(selectRatio * 100).toFixed(1)}%) - may include noise`);
|
||||
}
|
||||
}
|
||||
|
||||
// 实体过滤问题
|
||||
if (m.l2.entityFilterStats) {
|
||||
const ef = m.l2.entityFilterStats;
|
||||
if (ef.filtered === 0 && ef.before > 10) {
|
||||
issues.push(`No events filtered by entity - focus entities may be too broad or missing`);
|
||||
}
|
||||
if (ef.before > 0 && ef.filtered > ef.before * 0.8) {
|
||||
issues.push(`Too many events filtered (${ef.filtered}/${ef.before}) - focus may be too narrow`);
|
||||
}
|
||||
}
|
||||
|
||||
// 相似度问题
|
||||
if (m.l2.similarityDistribution && m.l2.similarityDistribution.min > 0 && m.l2.similarityDistribution.min < 0.5) {
|
||||
issues.push(`Low similarity events included (min=${m.l2.similarityDistribution.min})`);
|
||||
}
|
||||
|
||||
// 因果链问题
|
||||
if (m.l2.eventsSelected > 0 && m.l2.causalEventsCount === 0 && m.l2.byRecallType.direct === 0) {
|
||||
issues.push('No direct or causal events - query expansion may be inaccurate');
|
||||
}
|
||||
|
||||
// L0 atoms 问题
|
||||
if ((m.l0.atomsMatched || 0) === 0) {
|
||||
issues.push('L0 atoms not matched - may need to generate anchors');
|
||||
}
|
||||
|
||||
// Rerank 相关问题
|
||||
if (m.l3.rerankApplied) {
|
||||
if (m.l3.beforeRerank > 0 && m.l3.afterRerank > 0) {
|
||||
const filterRatio = 1 - (m.l3.afterRerank / m.l3.beforeRerank);
|
||||
if (filterRatio > 0.7) {
|
||||
issues.push(`High rerank filter ratio (${(filterRatio * 100).toFixed(0)}%) - many irrelevant chunks removed`);
|
||||
}
|
||||
}
|
||||
|
||||
if (m.l3.rerankScoreDistribution) {
|
||||
const rd = m.l3.rerankScoreDistribution;
|
||||
if (rd.max < 0.5) {
|
||||
issues.push(`Low rerank scores (max=${rd.max}) - query may be poorly matched`);
|
||||
}
|
||||
if (rd.mean < 0.3) {
|
||||
issues.push(`Very low average rerank score (mean=${rd.mean}) - context may be weak`);
|
||||
}
|
||||
}
|
||||
|
||||
if (m.l3.rerankTime > 2000) {
|
||||
issues.push(`Slow rerank (${m.l3.rerankTime}ms) - may affect response time`);
|
||||
}
|
||||
}
|
||||
|
||||
// 证据密度问题（基于 selected 的构成）
|
||||
if (m.l3.chunksSelected > 0 && m.l3.chunksSelectedByType) {
|
||||
const l1Real = m.l3.chunksSelectedByType.l1Real || 0;
|
||||
const density = l1Real / m.l3.chunksSelected;
|
||||
if (density < 0.3 && m.l3.chunksSelected > 10) {
|
||||
issues.push(`Low L1 chunk ratio in selected (${(density * 100).toFixed(0)}%) - may lack concrete evidence`);
|
||||
}
|
||||
}
|
||||
|
||||
// 预算问题
|
||||
if (m.budget.utilization > 90) {
|
||||
issues.push(`High budget utilization (${m.budget.utilization}%) - may be truncating content`);
|
||||
}
|
||||
|
||||
// 性能问题
|
||||
if (m.timing.total > 5000) {
|
||||
issues.push(`Slow recall (${m.timing.total}ms) - consider optimization`);
|
||||
}
|
||||
|
||||
return issues;
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -166,6 +166,7 @@ class StreamingGeneration {
|
||||
if (!source) throw new Error(`不支持的 api: ${opts.api}`);
|
||||
|
||||
const model = String(opts.model || '').trim();
|
||||
const msgCount = Array.isArray(messages) ? messages.length : null;
|
||||
|
||||
if (!model) {
|
||||
try { xbLog.error('streamingGeneration', 'missing model', null); } catch {}
|
||||
@@ -175,7 +176,6 @@ class StreamingGeneration {
|
||||
try {
|
||||
try {
|
||||
if (xbLog.isEnabled?.()) {
|
||||
const msgCount = Array.isArray(messages) ? messages.length : null;
|
||||
xbLog.info('streamingGeneration', `callAPI stream=${!!stream} api=${String(opts.api || '')} model=${model} messages=${msgCount ?? '-'}`);
|
||||
}
|
||||
} catch {}
|
||||
@@ -286,10 +286,34 @@ class StreamingGeneration {
|
||||
}
|
||||
|
||||
|
||||
const logSendRequestError = (err, streamMode) => {
|
||||
if (err?.name !== 'AbortError') {
|
||||
const safeApiUrl = String(cmdApiUrl || reverseProxy || oai_settings?.custom_url || '').trim();
|
||||
try {
|
||||
xbLog.error('streamingGeneration', 'sendRequest failed', {
|
||||
message: err?.message || String(err),
|
||||
name: err?.name,
|
||||
stream: !!streamMode,
|
||||
api: String(opts.api || ''),
|
||||
model,
|
||||
msgCount,
|
||||
apiurl: safeApiUrl,
|
||||
});
|
||||
} catch {}
|
||||
console.error('[xbgen:callAPI] sendRequest failed:', err);
|
||||
}
|
||||
};
|
||||
|
||||
if (stream) {
|
||||
const payload = ChatCompletionService.createRequestData(body);
|
||||
|
||||
const streamFactory = await ChatCompletionService.sendRequest(payload, false, abortSignal);
|
||||
let streamFactory;
|
||||
try {
|
||||
streamFactory = await ChatCompletionService.sendRequest(payload, false, abortSignal);
|
||||
} catch (err) {
|
||||
logSendRequestError(err, true);
|
||||
throw err;
|
||||
}
|
||||
|
||||
const generator = (typeof streamFactory === 'function') ? streamFactory() : streamFactory;
|
||||
|
||||
@@ -350,7 +374,13 @@ class StreamingGeneration {
|
||||
})();
|
||||
} else {
|
||||
const payload = ChatCompletionService.createRequestData(body);
|
||||
const extracted = await ChatCompletionService.sendRequest(payload, false, abortSignal);
|
||||
let extracted;
|
||||
try {
|
||||
extracted = await ChatCompletionService.sendRequest(payload, false, abortSignal);
|
||||
} catch (err) {
|
||||
logSendRequestError(err, false);
|
||||
throw err;
|
||||
}
|
||||
|
||||
let result = '';
|
||||
if (extracted && typeof extracted === 'object') {
|
||||
|
||||
Reference in New Issue
Block a user