Adjust summary prompt labels and budgets

modules/story-summary/data/store.js

@@ -5,10 +5,11 @@ import { getContext, saveMetadataDebounced } from "../../../../../../extensions.
 import { chat_metadata } from "../../../../../../../script.js";
 import { EXT_ID } from "../../../core/constants.js";
 import { xbLog } from "../../../core/debug-core.js";
-import { clearEventVectors, deleteEventVectorsByIds } from "../vector/chunk-store.js";
+import { clearEventVectors, deleteEventVectorsByIds } from "../vector/storage/chunk-store.js";
-import { clearEventTextIndex } from '../vector/text-search.js';
+import { clearEventTextIndex } from '../vector/retrieval/text-search.js';
 
 const MODULE_ID = 'summaryStore';
+const FACTS_LIMIT_PER_SUBJECT = 10;
 
 // ═══════════════════════════════════════════════════════════════════════════
 // 基础存取
@@ -125,32 +126,26 @@ function getNextFactId(existingFacts) {
 export function mergeFacts(existingFacts, updates, floor) {
     const map = new Map();
 
-    // 加载现有 facts
     for (const f of existingFacts || []) {
         if (!f.retracted) {
             map.set(factKey(f), f);
         }
     }
 
-    // 获取下一个 ID
     let nextId = getNextFactId(existingFacts);
 
-    // 应用更新
     for (const u of updates || []) {
         if (!u.s || !u.p) continue;
 
         const key = factKey(u);
 
-        // 删除操作
         if (u.retracted === true) {
             map.delete(key);
             continue;
         }
 
-        // 无 o 则跳过
         if (!u.o || !String(u.o).trim()) continue;
 
-        // 覆盖或新增
         const existing = map.get(key);
         const newFact = {
             id: existing?.id || `f-${nextId++}`,
@@ -158,14 +153,13 @@ export function mergeFacts(existingFacts, updates, floor) {
             p: u.p.trim(),
             o: String(u.o).trim(),
             since: floor,
+            _isState: existing?._isState ?? !!u.isState,
         };
 
-        // 关系类保留 trend
         if (isRelationFact(newFact) && u.trend) {
             newFact.trend = u.trend;
         }
 
-        // 保留原始 _addedAt（如果是更新）
         if (existing?._addedAt != null) {
             newFact._addedAt = existing._addedAt;
         } else {
@@ -175,9 +169,28 @@ export function mergeFacts(existingFacts, updates, floor) {
         map.set(key, newFact);
     }
 
-    return Array.from(map.values());
+    const factsBySubject = new Map();
+    for (const f of map.values()) {
+        if (f._isState) continue;
+        const arr = factsBySubject.get(f.s) || [];
+        arr.push(f);
+        factsBySubject.set(f.s, arr);
+    }
+
+    const toRemove = new Set();
+    for (const arr of factsBySubject.values()) {
+        if (arr.length > FACTS_LIMIT_PER_SUBJECT) {
+            arr.sort((a, b) => (a._addedAt || 0) - (b._addedAt || 0));
+            for (let i = 0; i < arr.length - FACTS_LIMIT_PER_SUBJECT; i++) {
+                toRemove.add(factKey(arr[i]));
+            }
+        }
+    }
+
+    return Array.from(map.values()).filter(f => !toRemove.has(factKey(f)));
 }
 
+
 // ═══════════════════════════════════════════════════════════════════════════
 // 旧数据迁移
 // ═══════════════════════════════════════════════════════════════════════════

modules/story-summary/generate/generator.js

@@ -32,7 +32,6 @@ function sanitizeFacts(parsed) {
 
         if (!s || !pRaw) continue;
 
-        // 删除操作
         if (item.retracted === true) {
             ok.push({ s, p: pRaw, retracted: true });
             continue;
@@ -43,11 +42,15 @@ function sanitizeFacts(parsed) {
 
         const relP = normalizeRelationPredicate(pRaw);
         const isRel = !!relP;
-        const fact = { s, p: isRel ? relP : pRaw, o };
+        const fact = {
+            s,
+            p: isRel ? relP : pRaw,
+            o,
+            isState: !!item.isState,
+        };
 
-        // 关系类保留 trend
         if (isRel && item.trend) {
-            const validTrends = ['破裂', '厌恶', '反感', '陌生', '投缘', '亲密', '交融'];
+            const validTrends = ['??', '??', '??', '??', '??', '??', '??'];
             if (validTrends.includes(item.trend)) {
                 fact.trend = item.trend;
             }
@@ -59,6 +62,7 @@ function sanitizeFacts(parsed) {
     parsed.factUpdates = ok;
 }
 
+
 // ═══════════════════════════════════════════════════════════════════════════
 // causedBy 清洗（事件因果边）
 // ═══════════════════════════════════════════════════════════════════════════

modules/story-summary/generate/llm.js

@@ -100,14 +100,19 @@ Acknowledged. Now reviewing the incremental summarization specifications:
 ├─ progress: 0.0 to 1.0
 └─ newMoment: 仅记录本次新增的关键时刻
 
-[Fact Tracking - SPO Triples]
+[Fact Tracking - SPO ???]
-├─ s: 主体（角色名/物品名）
+?? ??: ?? & ???????
-├─ p: 谓词（属性名）
+?? ??: ??????????????????
-│   - 关系类只允许：对X的看法 / 与X的关系
+?? SPO ??:
-├─ o: 值（当前状态）
+?   s: ??????/????
-├─ trend: 仅关系类填写
+?   p: ??????????????
-├─ retracted: 删除标记
+?   o: ???
-└─ s+p 为键，相同键会覆盖旧值
+?? KV ??: s+p ??????????
+?? isState ????????:
+?   true  = ????????????/??/??/???
+?   false = ??????????????
+?? trend: ?????????/??/??/??/??/??/???
+?? retracted: true ???????
 
 Ready to process incremental summary requests with strict deduplication.`,
 
@@ -177,26 +182,28 @@ Before generating, observe the USER and analyze carefully:
   "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"}
   ],
-  "factUpdates": [
+    "factUpdates": [
     {
-      "s": "主体（角色名/物品名）",
+      "s": "主体",
-      "p": "谓词（属性名/对X的看法）",
+      "p": "谓词（复用已有谓词，避免同义词）",
       "o": "当前值",
-      "trend": "破裂|厌恶|反感|陌生|投缘|亲密|交融",
+      "isState": true/false,
-      "retracted": false
+      "trend": "仅关系类：破裂|厌恶|反感|陌生|投缘|亲密|交融"
     }
   ]
 }
+
+
 \`\`\`
 
 ## 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
-- 状态类：s=角色名, p=属性(生死/位置/状态等), o=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：s=角色A, p="对B的看法" 或 p="与B的关系"，trend 仅限关系类
+- 关系类: p="对X的看法"，trend 必填
-- 删除：设置 retracted: true（不需要填 o）
+- 删除: 设置 retracted: true
-- 只输出有变化的条目
+- 谓词规范化: 复用已有谓词，不要发明同义词
-- 硬约束才记录，避免叙事化，确保少、硬、稳定
+- 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
@@ -267,9 +274,11 @@ function waitForStreamingComplete(sessionId, streamingMod, timeout = 120000) {
 
 function formatFactsForLLM(facts) {
     if (!facts?.length) {
-        return '（空白，尚无事实记录）';
+        return { text: '（空白，尚无事实记录）', predicates: [] };
     }
 
+    const predicates = [...new Set(facts.map(f => f.p).filter(Boolean))];
+
     const lines = facts.map(f => {
         if (f.trend) {
             return `- ${f.s} | ${f.p} | ${f.o} [${f.trend}]`;
@@ -277,11 +286,18 @@ function formatFactsForLLM(facts) {
         return `- ${f.s} | ${f.p} | ${f.o}`;
     });
 
-    return lines.join('\n') || '（空白，尚无事实记录）';
+    return {
+        text: lines.join('\n') || '（空白，尚无事实记录）',
+        predicates,
+    };
 }
 
 function buildSummaryMessages(existingSummary, existingFacts, newHistoryText, historyRange, nextEventId, existingEventCount) {
-    const factsText = formatFactsForLLM(existingFacts);
+    const { text: factsText, predicates } = formatFactsForLLM(existingFacts);
+
+    const predicatesHint = predicates.length > 0
+        ? `\n\n<\u5df2\u6709\u8c13\u8bcd\uff0c\u8bf7\u590d\u7528>\n${predicates.join('\u3001')}\n</\u5df2\u6709\u8c13\u8bcd\uff0c\u8bf7\u590d\u7528>`
+        : '';
 
     const jsonFormat = LLM_PROMPT_CONFIG.userJsonFormat
         .replace(/\{nextEventId\}/g, String(nextEventId));
@@ -293,9 +309,9 @@ function buildSummaryMessages(existingSummary, existingFacts, newHistoryText, hi
         { role: 'system', content: LLM_PROMPT_CONFIG.topSystem },
         { role: 'assistant', content: LLM_PROMPT_CONFIG.assistantDoc },
         { role: 'assistant', content: LLM_PROMPT_CONFIG.assistantAskSummary },
-        { role: 'user', content: `<已有总结状态>\n${existingSummary}\n</已有总结状态>\n\n<当前事实图谱>\n${factsText}\n</当前事实图谱>` },
+        { role: 'user', content: `<\u5df2\u6709\u603b\u7ed3\u72b6\u6001>\n${existingSummary}\n</\u5df2\u6709\u603b\u7ed3\u72b6\u6001>\n\n<\u5f53\u524d\u4e8b\u5b9e\u56fe\u8c31>\n${factsText}\n</\u5f53\u524d\u4e8b\u5b9e\u56fe\u8c31>${predicatesHint}` },
         { role: 'assistant', content: LLM_PROMPT_CONFIG.assistantAskContent },
-        { role: 'user', content: `<新对话内容>（${historyRange}）\n${newHistoryText}\n</新对话内容>` }
+        { role: 'user', content: `<\u65b0\u5bf9\u8bdd\u5185\u5bb9>\uff08${historyRange}\uff09\n${newHistoryText}\n</\u65b0\u5bf9\u8bdd\u5185\u5bb9>` }
     ];
 
     const bottomMessages = [
@@ -311,6 +327,7 @@ function buildSummaryMessages(existingSummary, existingFacts, newHistoryText, hi
     };
 }
 
+
 // ═══════════════════════════════════════════════════════════════════════════
 // JSON 解析
 // ═══════════════════════════════════════════════════════════════════════════
@@ -415,4 +432,4 @@ export async function generateSummary(options) {
     console.groupEnd();
 
     return rawOutput;
-}
+}

modules/story-summary/generate/prompt.js

@@ -1,6 +1,6 @@
-// ═══════════════════════════════════════════════════════════════════════════
+// ═══════════════════════════════════════════════════════════════════════════
 // Story Summary - Prompt Injection (Final Clean Version)
-// - 仅负责“构建注入文本”，不负责写入 extension_prompts
+// - 仅负责"构建注入文本"，不负责写入 extension_prompts
 // - 注入发生在 story-summary.js：GENERATION_STARTED 时写入 extension_prompts（IN_CHAT + depth）
 // ═══════════════════════════════════════════════════════════════════════════
 
@@ -8,8 +8,8 @@ import { getContext } from "../../../../../../extensions.js";
 import { xbLog } from "../../../core/debug-core.js";
 import { getSummaryStore, getFacts, isRelationFact } from "../data/store.js";
 import { getVectorConfig, getSummaryPanelConfig, getSettings } from "../data/config.js";
-import { recallMemory, buildQueryText } from "../vector/recall.js";
+import { recallMemory, buildQueryText } from "../vector/retrieval/recall.js";
-import { getChunksByFloors, getAllChunkVectors, getAllEventVectors, getMeta } from "../vector/chunk-store.js";
+import { getChunksByFloors, getAllChunkVectors, getAllEventVectors, getMeta } from "../vector/storage/chunk-store.js";
 
 const MODULE_ID = "summaryPrompt";
 
@@ -32,6 +32,7 @@ function canNotifyRecallFail() {
 // ─────────────────────────────────────────────────────────────────────────────
 
 const MAIN_BUDGET_MAX = 10000;   // 主装配预算（世界/事件/远期/弧光）
+const ORPHAN_MAX = 2500;         // 远期上限
 const RECENT_ORPHAN_MAX = 5000;  // [待整理] 独立预算
 const TOTAL_BUDGET_MAX = 15000;  // 总预算（用于日志显示）
 const L3_MAX = 2000;
@@ -85,6 +86,49 @@ function cleanSummary(summary) {
         .trim();
 }
 
+// ─────────────────────────────────────────────────────────────────────────────
+// 上下文配对工具函数
+// ─────────────────────────────────────────────────────────────────────────────
+
+/**
+ * 获取chunk的配对楼层
+ * USER楼层 → 下一楼(AI回复)
+ * AI楼层   → 上一楼(USER发言)
+ */
+function getContextFloor(chunk) {
+    if (chunk.isL0) return -1; // L0虚拟chunk不需要配对
+    return chunk.isUser ? chunk.floor + 1 : chunk.floor - 1;
+}
+
+/**
+ * 从候选chunks中选择最佳配对
+ * 策略：优先选择相反角色的第一个chunk
+ */
+function pickContextChunk(candidates, mainChunk) {
+    if (!candidates?.length) return null;
+    const targetIsUser = !mainChunk.isUser;
+    // 优先相反角色
+    const opposite = candidates.find(c => c.isUser === targetIsUser);
+    if (opposite) return opposite;
+    // 否则选第一个
+    return candidates[0];
+}
+/**
+ * 格式化配对chunk（完整显示，带缩进和方向符号）
+ */
+function formatContextChunkLine(chunk, isAbove) {
+    const { name1, name2 } = getContext();
+    const speaker = chunk.isUser ? (name1 || "用户") : (chunk.speaker || name2 || "角色");
+    const text = String(chunk.text || "").trim();
+    const symbol = isAbove ? "┌" : "└";
+    return `  ${symbol} #${chunk.floor + 1} [${speaker}] ${text}`;
+}
+
+/**
+ * 格式化配对chunk（缩进，简短摘要）
+ */
+
+
 // ─────────────────────────────────────────────────────────────────────────────
 // 系统前导与后缀
 // ─────────────────────────────────────────────────────────────────────────────
@@ -131,7 +175,7 @@ function formatArcLine(a) {
         .filter(Boolean);
 
     if (moments.length) {
-        return `- ${a.name}：${moments.join(" → ")}（当前：${a.trajectory}）`;
+        return `- ${a.name}：${moments.join(" → ")}`;
     }
     return `- ${a.name}：${a.trajectory}`;
 }
@@ -150,7 +194,31 @@ function formatChunkFullLine(c) {
     return `› #${c.floor + 1} [${speaker}] ${String(c.text || "").trim()}`;
 }
 
-// 因果事件格式（仅作为“前因线索”展示，仍保留楼层提示）
+/**
+ * 格式化chunk及其配对上下文
+ * 返回数组：[配对行(如果在前), 主chunk行, 配对行(如果在后)]
+ */
+function formatChunkWithContext(mainChunk, contextChunk) {
+    const lines = [];
+    const mainLine = formatChunkFullLine(mainChunk);
+
+    if (!contextChunk) {
+        lines.push(mainLine);
+        return lines;
+    }
+
+    if (contextChunk.floor < mainChunk.floor) {
+        lines.push(formatContextChunkLine(contextChunk, true));
+        lines.push(mainLine);
+    } else {
+        lines.push(mainLine);
+        lines.push(formatContextChunkLine(contextChunk, false));
+    }
+
+    return lines;
+}
+
+// 因果事件格式（仅作为"前因线索"展示，仍保留楼层提示）
 function formatCausalEventLine(causalItem, causalById) {
     const ev = causalItem?.event || {};
     const depth = Math.max(1, Math.min(9, causalItem?._causalDepth || 1));
@@ -172,9 +240,8 @@ function formatCausalEventLine(causalItem, causalById) {
     const evidence = causalItem._evidenceChunk;
     if (evidence) {
         const speaker = evidence.speaker || "角色";
-        const preview = String(evidence.text || "");
+        const text = String(evidence.text || "").trim();
-        const clip = preview.length > 60 ? preview.slice(0, 60) + "..." : preview;
+        lines.push(`${indent}  › #${evidence.floor + 1} [${speaker}] ${text}`);
-        lines.push(`${indent}  › #${evidence.floor + 1} [${speaker}] ${clip}`);
     }
 
     return lines.join("\n");
@@ -216,11 +283,13 @@ function formatInjectionLog(stats, details, recentOrphanStats = null) {
     const l1OrphanCount = (stats.orphans.injected || 0) - l0OrphanCount;
     lines.push(`  [3] 远期片段 (已总结范围)`);
     lines.push(`      选入: ${stats.orphans.injected} 条 (L0: ${l0OrphanCount}, L1: ${l1OrphanCount}) | 消耗: ${stats.orphans.tokens} tokens`);
+    lines.push(`      配对: ${stats.orphans.contextPairs || 0} 条`);
     lines.push('');
 
     // [4] 待整理
     lines.push(`  [4] 待整理 (独立预算 5000)`);
     lines.push(`      选入: ${recentOrphanStats?.injected || 0} 条 | 消耗: ${recentOrphanStats?.tokens || 0} tokens`);
+    lines.push(`      配对: ${recentOrphanStats?.contextPairs || 0} 条`);
     lines.push(`      楼层: ${recentOrphanStats?.floorRange || 'N/A'}`);
     lines.push('');
 
@@ -248,7 +317,7 @@ function formatInjectionLog(stats, details, recentOrphanStats = null) {
     return lines.join('\n');
 }
 
-// 重写事件文本里的序号前缀：把 “{idx}. ” 或 “{idx}.【...】” 的 idx 替换
+// 重写事件文本里的序号前缀：把 "{idx}. " 或 "{idx}.【...】" 的 idx 替换
 function renumberEventText(text, newIndex) {
     const s = String(text || "");
     // 匹配行首：  "12."  或  "12.【"
@@ -325,11 +394,12 @@ export function buildNonVectorPromptText() {
     return text;
 }
 
-// ─────────────────────────────────────────────────────────────
+// ─────────────────────────────────────────────────────────────────────────────
 // 向量模式：预算装配（世界 → 事件(带证据) → 碎片 → 弧光）
-// ─────────────────────────────────────────────────────────────
+// ─────────────────────────────────────────────────────────────────────────────
 
 async function buildVectorPrompt(store, recallResult, causalById, queryEntities = [], meta = null) {
+    const { chatId } = getContext();
     const data = store.json || {};
     const total = { used: 0, max: MAIN_BUDGET_MAX };
 
@@ -351,13 +421,14 @@ async function buildVectorPrompt(store, recallResult, causalById, queryEntities
         arcs: { count: 0, tokens: 0 },
         events: { selected: 0, tokens: 0 },
         evidence: { attached: 0, tokens: 0 },
-        orphans: { injected: 0, tokens: 0 },
+        orphans: { injected: 0, tokens: 0, l0Count: 0, contextPairs: 0 },
     };
 
     const recentOrphanStats = {
         injected: 0,
         tokens: 0,
         floorRange: "N/A",
+        contextPairs: 0,
     };
     const details = {
         eventList: [],
@@ -473,14 +544,14 @@ async function buildVectorPrompt(store, recallResult, causalById, queryEntities
 
         const bestChunk = pickBestChunkForEvent(e.event);
 
-        // 先尝试“带证据”
+        // 先尝试"带证据"
         // idx 先占位写 0，后面统一按时间线重排后再改号
         let text = formatEventWithEvidence(e, 0, bestChunk);
         let cost = estimateTokens(text);
         let hasEvidence = !!bestChunk;
         let chosenChunk = bestChunk || null;
 
-        // 塞不下就退化成“不带证据”
+        // 塞不下就退化成"不带证据"
         if (total.used + cost > total.max) {
             text = formatEventWithEvidence(e, 0, null);
             cost = estimateTokens(text);
@@ -549,33 +620,90 @@ async function buildVectorPrompt(store, recallResult, causalById, queryEntities
     assembled.events.similar = selectedSimilarTexts;
 
     // ═══════════════════════════════════════════════════════════════════
-    // [优先级 4] 远期片段（已总结范围的 orphan chunks）
+    // [优先级 4] 远期片段（已总结范围的 orphan chunks）- 带上下文配对
     // ═══════════════════════════════════════════════════════════════════
     const lastSummarized = store.lastSummarizedMesId ?? -1;
     const lastChunkFloor = meta?.lastChunkFloor ?? -1;
     const keepVisible = store.keepVisibleCount ?? 3;
 
-    if (chunks.length && total.used < total.max) {
+    // 收集需要配对的楼层
-        const orphans = chunks
+    const orphanContextFloors = new Set();
-            .filter(c => !usedChunkIds.has(c.chunkId))
+    const orphanCandidates = chunks
-            .filter(c => c.floor <= lastSummarized)
+        .filter(c => !usedChunkIds.has(c.chunkId))
+        .filter(c => c.floor <= lastSummarized);
+
+    for (const c of orphanCandidates) {
+        if (c.isL0) continue;
+        const pairFloor = getContextFloor(c);
+        if (pairFloor >= 0) orphanContextFloors.add(pairFloor);
+    }
+
+    // 批量获取配对楼层的chunks
+    let contextChunksByFloor = new Map();
+    if (chatId && orphanContextFloors.size > 0) {
+        try {
+            const contextChunks = await getChunksByFloors(chatId, Array.from(orphanContextFloors));
+            for (const pc of contextChunks) {
+                if (!contextChunksByFloor.has(pc.floor)) {
+                    contextChunksByFloor.set(pc.floor, []);
+                }
+                contextChunksByFloor.get(pc.floor).push(pc);
+            }
+        } catch (e) {
+            xbLog.warn(MODULE_ID, "获取配对chunks失败", e);
+        }
+    }
+
+    if (orphanCandidates.length && total.used < total.max) {
+        const orphans = orphanCandidates
             .sort((a, b) => (a.floor - b.floor) || ((a.chunkIdx ?? 0) - (b.chunkIdx ?? 0)));
 
-        const l1Budget = { used: 0, max: total.max - total.used };
+        const l1Budget = { used: 0, max: Math.min(ORPHAN_MAX, total.max - total.used) };
+        let l0Count = 0;
+        let contextPairsCount = 0;
 
         for (const c of orphans) {
-            const line = formatChunkFullLine(c);
+            // L0 不需要配对
-            if (!pushWithBudget(assembled.orphans.lines, line, l1Budget)) break;
+            if (c.isL0) {
+                const line = formatChunkFullLine(c);
+                if (!pushWithBudget(assembled.orphans.lines, line, l1Budget)) break;
+                injectionStats.orphans.injected++;
+                l0Count++;
+                continue;
+            }
+
+            // 获取配对chunk
+            const pairFloor = getContextFloor(c);
+            const candidates = contextChunksByFloor.get(pairFloor) || [];
+            const contextChunk = pickContextChunk(candidates, c);
+
+            // 格式化（带配对）
+            const formattedLines = formatChunkWithContext(c, contextChunk);
+
+            // 尝试添加所有行
+            let allAdded = true;
+            for (const line of formattedLines) {
+                if (!pushWithBudget(assembled.orphans.lines, line, l1Budget)) {
+                    allAdded = false;
+                    break;
+                }
+            }
+
+            if (!allAdded) break;
+
             injectionStats.orphans.injected++;
+            if (contextChunk) contextPairsCount++;
         }
 
         assembled.orphans.tokens = l1Budget.used;
         total.used += l1Budget.used;
         injectionStats.orphans.tokens = l1Budget.used;
+        injectionStats.orphans.l0Count = l0Count;
+        injectionStats.orphans.contextPairs = contextPairsCount;
     }
 
     // ═══════════════════════════════════════════════════════════════════
-    // [独立预算] 待整理（未总结范围，独立 5000）
+    // [独立预算] 待整理（未总结范围，独立 5000）- 带上下文配对
     // ═══════════════════════════════════════════════════════════════════
 
     // 近期范围：(lastSummarized, lastChunkFloor - keepVisible]
@@ -583,55 +711,113 @@ async function buildVectorPrompt(store, recallResult, causalById, queryEntities
     const recentEnd = lastChunkFloor - keepVisible;
 
     if (chunks.length && recentEnd >= recentStart) {
-        const recentOrphans = chunks
+        const recentOrphanCandidates = chunks
             .filter(c => !usedChunkIds.has(c.chunkId))
-            .filter(c => c.floor >= recentStart && c.floor <= recentEnd)
+            .filter(c => c.floor >= recentStart && c.floor <= recentEnd);
+
+        // 收集近期范围需要配对的楼层
+        const recentContextFloors = new Set();
+        for (const c of recentOrphanCandidates) {
+            if (c.isL0) continue;
+            const pairFloor = getContextFloor(c);
+            if (pairFloor >= 0) recentContextFloors.add(pairFloor);
+        }
+
+        // 批量获取（复用已有的 or 新获取）
+        let recentContextChunksByFloor = new Map();
+        if (chatId && recentContextFloors.size > 0) {
+            // 过滤掉已经获取过的
+            const newFloors = Array.from(recentContextFloors).filter(f => !contextChunksByFloor.has(f));
+            if (newFloors.length > 0) {
+                try {
+                    const newContextChunks = await getChunksByFloors(chatId, newFloors);
+                    for (const pc of newContextChunks) {
+                        if (!contextChunksByFloor.has(pc.floor)) {
+                            contextChunksByFloor.set(pc.floor, []);
+                        }
+                        contextChunksByFloor.get(pc.floor).push(pc);
+                    }
+                } catch (e) {
+                    xbLog.warn(MODULE_ID, "获取近期配对chunks失败", e);
+                }
+            }
+            recentContextChunksByFloor = contextChunksByFloor;
+        }
+
+        const recentOrphans = recentOrphanCandidates
             .sort((a, b) => (a.floor - b.floor) || ((a.chunkIdx ?? 0) - (b.chunkIdx ?? 0)));
 
         const recentBudget = { used: 0, max: RECENT_ORPHAN_MAX };
+        let recentContextPairsCount = 0;
 
         for (const c of recentOrphans) {
-            const line = formatChunkFullLine(c);
+            // L0 不需要配对
-            if (!pushWithBudget(assembled.recentOrphans.lines, line, recentBudget)) break;
+            if (c.isL0) {
+                const line = formatChunkFullLine(c);
+                if (!pushWithBudget(assembled.recentOrphans.lines, line, recentBudget)) break;
+                recentOrphanStats.injected++;
+                continue;
+            }
+
+            // 获取配对chunk
+            const pairFloor = getContextFloor(c);
+            const candidates = recentContextChunksByFloor.get(pairFloor) || [];
+            const contextChunk = pickContextChunk(candidates, c);
+
+            // 格式化（带配对）
+            const formattedLines = formatChunkWithContext(c, contextChunk);
+
+            // 尝试添加所有行
+            let allAdded = true;
+            for (const line of formattedLines) {
+                if (!pushWithBudget(assembled.recentOrphans.lines, line, recentBudget)) {
+                    allAdded = false;
+                    break;
+                }
+            }
+
+            if (!allAdded) break;
+
             recentOrphanStats.injected++;
+            if (contextChunk) recentContextPairsCount++;
         }
 
         assembled.recentOrphans.tokens = recentBudget.used;
         recentOrphanStats.tokens = recentBudget.used;
         recentOrphanStats.floorRange = `${recentStart + 1}~${recentEnd + 1}楼`;
+        recentOrphanStats.contextPairs = recentContextPairsCount;
     }
 
     // ═══════════════════════════════════════════════════════════════════
-    // ═══════════════════════════════════════════════════════════════════════
+    // 按注入顺序拼接 sections
-// 按注入顺序拼接 sections
+    // ═══════════════════════════════════════════════════════════════════
-// ═══════════════════════════════════════════════════════════════════════
+    const sections = [];
-const sections = [];
+    // 1. 世界约束 → 定了的事
-// 1. 世界约束 → 定了的事
+    if (assembled.facts.lines.length) {
-if (assembled.facts.lines.length) {
+        sections.push(`[定了的事] 已确立的事实\n${assembled.facts.lines.join("\n")}`);
-    sections.push(`[定了的事] 已确立的事实\n${assembled.facts.lines.join("\n")}`);
+    }
-}
+    // 2. 核心经历 → 印象深的事
-// 2. 核心经历 → 印象深的事
+    if (assembled.events.direct.length) {
-if (assembled.events.direct.length) {
+        sections.push(`[印象深的事] 记得很清楚\n\n${assembled.events.direct.join("\n\n")}`);
-    sections.push(`[印象深的事] 记得很清楚\n\n${assembled.events.direct.join("\n\n")}`);
+    }
-}
+    // 3. 过往背景 → 好像有关的事
-// 3. 过往背景 → 好像有关的事
+    if (assembled.events.similar.length) {
-if (assembled.events.similar.length) {
+        sections.push(`[好像有关的事] 听说过或有点模糊\n\n${assembled.events.similar.join("\n\n")}`);
-    sections.push(`[好像有关的事] 听说过或有点模糊\n\n${assembled.events.similar.join("\n\n")}`);
+    }
-}
+    // 4. 远期片段 → 更早以前
-// 4. 远期片段 → 更早以前
+    if (assembled.orphans.lines.length) {
-if (assembled.orphans.lines.length) {
+        sections.push(`[更早以前] 记忆里残留的老画面\n${assembled.orphans.lines.join("\n")}`);
-    sections.push(`[更早以前] 记忆里残留的老画面\n${assembled.orphans.lines.join("\n")}`);
+    }
-}
+    // 5. 待整理 → 近期
-// 5. 待整理 → 刚发生的
+    if (assembled.recentOrphans.lines.length) {
-if (assembled.recentOrphans.lines.length) {
+        sections.push(`[近期] 清晰但还没整理\n${assembled.recentOrphans.lines.join("\n")}`);
-    sections.push(`[刚发生的] 还没来得及想明白\n${assembled.recentOrphans.lines.join("\n")}`);
+    }
-}
+    // 6. 人物弧光 → 这些人
-// 6. 人物弧光 → 这些人
+    if (assembled.arcs.lines.length) {
-if (assembled.arcs.lines.length) {
+        sections.push(`[这些人] 他们的弧光\n${assembled.arcs.lines.join("\n")}`);
-    sections.push(`[这些人] 他们现在怎样了\n${assembled.arcs.lines.join("\n")}`);
+    }
-}
 
-if (!sections.length) {
+    if (!sections.length) {
         return { promptText: "", injectionLogText: "", injectionStats };
     }
 
@@ -846,3 +1032,4 @@ export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
 
     return { text: finalText, logText: (recallResult.logText || "") + (injectionLogText || "") };
 }
+

modules/story-summary/story-summary.js

@@ -54,7 +54,7 @@ import {
     fetchOnlineModels,
     isLocalModelLoaded,
     DEFAULT_LOCAL_MODEL,
-} from "./vector/embedder.js";
+} from "./vector/utils/embedder.js";
 
 import {
     getMeta,
@@ -66,7 +66,7 @@ import {
     saveChunks,
     saveChunkVectors,
     getStorageStats,
-} from "./vector/chunk-store.js";
+} from "./vector/storage/chunk-store.js";
 
 import {
     buildIncrementalChunks,
@@ -75,12 +75,12 @@ import {
     syncOnMessageDeleted,
     syncOnMessageSwiped,
     syncOnMessageReceived,
-} from "./vector/chunk-builder.js";
+} from "./vector/pipeline/chunk-builder.js";
-import { initStateIntegration, rebuildStateVectors } from "./vector/state-integration.js";
+import { initStateIntegration, rebuildStateVectors } from "./vector/pipeline/state-integration.js";
-import { clearStateVectors, getStateAtomsCount, getStateVectorsCount } from "./vector/state-store.js";
+import { clearStateVectors, getStateAtomsCount, getStateVectorsCount } from "./vector/storage/state-store.js";
 
 // vector io
-import { exportVectors, importVectors } from "./vector/vector-io.js";
+import { exportVectors, importVectors } from "./vector/storage/vector-io.js";
 
 // ═══════════════════════════════════════════════════════════════════════════
 // 常量

modules/story-summary/vector/pipeline/chunk-builder.js

@@ -3,7 +3,7 @@
 // 标准 RAG chunking: ~200 tokens per chunk
 // ═══════════════════════════════════════════════════════════════════════════
 
-import { getContext } from '../../../../../../extensions.js';
+import { getContext } from '../../../../../../../extensions.js';
 import {
     getMeta,
     updateMeta,
@@ -15,10 +15,10 @@ import {
     makeChunkId,
     hashText,
     CHUNK_MAX_TOKENS,
-} from './chunk-store.js';
+} from '../storage/chunk-store.js';
-import { embed, getEngineFingerprint } from './embedder.js';
+import { embed, getEngineFingerprint } from '../utils/embedder.js';
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
-import { filterText } from './text-filter.js';
+import { filterText } from '../utils/text-filter.js';
 
 const MODULE_ID = 'chunk-builder';
 
@@ -339,7 +339,7 @@ export async function syncOnMessageReceived(chatId, lastFloor, message, vectorCo
 
     // 本地模型未加载时跳过（避免意外触发下载或报错）
     if (vectorConfig.engine === "local") {
-        const { isLocalModelLoaded, DEFAULT_LOCAL_MODEL } = await import("./embedder.js");
+        const { isLocalModelLoaded, DEFAULT_LOCAL_MODEL } = await import("../utils/embedder.js");
         const modelId = vectorConfig.local?.modelId || DEFAULT_LOCAL_MODEL;
         if (!isLocalModelLoaded(modelId)) return;
     }

modules/story-summary/vector/pipeline/state-integration.js

@@ -3,8 +3,8 @@
 // 事件监听 + 回滚钩子注册
 // ═══════════════════════════════════════════════════════════════════════════
 
-import { getContext } from '../../../../../../extensions.js';
+import { getContext } from '../../../../../../../extensions.js';
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
 import {
     saveStateAtoms,
     saveStateVectors,
@@ -12,9 +12,9 @@ import {
     deleteStateVectorsFromFloor,
     getStateAtoms,
     clearStateVectors,
-} from './state-store.js';
+} from '../storage/state-store.js';
-import { embed, getEngineFingerprint } from './embedder.js';
+import { embed, getEngineFingerprint } from '../utils/embedder.js';
-import { getVectorConfig } from '../data/config.js';
+import { getVectorConfig } from '../../data/config.js';
 
 const MODULE_ID = 'state-integration';
 

modules/story-summary/vector/pipeline/state-recall.js

@@ -3,11 +3,11 @@
 // L0 语义锚点召回 + floor bonus + 虚拟 chunk 转换
 // ═══════════════════════════════════════════════════════════════════════════
 
-import { getContext } from '../../../../../../extensions.js';
+import { getContext } from '../../../../../../../extensions.js';
-import { getAllStateVectors, getStateAtoms } from './state-store.js';
+import { getAllStateVectors, getStateAtoms } from '../storage/state-store.js';
-import { getMeta } from './chunk-store.js';
+import { getMeta } from '../storage/chunk-store.js';
-import { getEngineFingerprint } from './embedder.js';
+import { getEngineFingerprint } from '../utils/embedder.js';
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
 
 const MODULE_ID = 'state-recall';
 

modules/story-summary/vector/retrieval/recall.js

@@ -1,4 +1,4 @@
-// Story Summary - Recall Engine
+// Story Summary - Recall Engine
 // L1 chunk + L2 event 召回
 // - 全量向量打分
 // - 实体权重归一化分配
@@ -8,19 +8,19 @@
 // - MMR 去重（融合后执行）
 // - floor 稀疏去重
 
-import { getAllEventVectors, getAllChunkVectors, getChunksByFloors, getMeta } from './chunk-store.js';
+import { getAllChunks, getAllEventVectors, getAllChunkVectors, getChunksByFloors, getMeta } from '../storage/chunk-store.js';
-import { embed, getEngineFingerprint } from './embedder.js';
+import { embed, getEngineFingerprint } from '../utils/embedder.js';
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
-import { getContext } from '../../../../../../extensions.js';
+import { getContext } from '../../../../../../../extensions.js';
-import { getSummaryStore, getFacts, getNewCharacters, isRelationFact } from '../data/store.js';
+import { getSummaryStore, getFacts, getNewCharacters, isRelationFact } from '../../data/store.js';
-import { filterText } from './text-filter.js';
+import { filterText } from '../utils/text-filter.js';
 import {
     searchStateAtoms,
     buildL0FloorBonus,
     stateToVirtualChunks,
     mergeAndSparsify,
-} from './state-recall.js';
+} from '../pipeline/state-recall.js';
-import { ensureEventTextIndex, searchEventsByText } from './text-search.js';
+import { ensureEventTextIndex, searchEventsByText, ensureChunkTextIndex, searchChunksByText } from './text-search.js';
 import {
     extractRareTerms,
     extractNounsFromFactsO,
@@ -29,10 +29,8 @@ import {
 const MODULE_ID = 'recall';
 
 const CONFIG = {
-    QUERY_MSG_COUNT: 5,
+    QUERY_MSG_COUNT: 3,
-    QUERY_DECAY_BETA: 0.7,
+    QUERY_DECAY_BETA: 0.6,
-    QUERY_MAX_CHARS: 600,
-    QUERY_CONTEXT_CHARS: 240,
 
     CAUSAL_CHAIN_MAX_DEPTH: 10,
     CAUSAL_INJECT_MAX: 30,
@@ -216,11 +214,26 @@ function extractRelationTarget(p) {
     return '';
 }
 
-function buildExpDecayWeights(n, beta) {
+function buildContentAwareWeights(segments, beta = 0.6) {
+    const n = segments.length;
+    if (n === 0) return [];
+    if (n === 1) return [1.0];
+
     const last = n - 1;
-    const w = Array.from({ length: n }, (_, i) => Math.exp(beta * (i - last)));
+    const SHORT_THRESHOLD = 15;
-    const sum = w.reduce((a, b) => a + b, 0) || 1;
+    const raw = [];
-    return w.map(x => x / sum);
+
+    for (let i = 0; i < n; i++) {
+        const posWeight = Math.exp(beta * (i - last));
+        const len = String(segments[i] || '').replace(/\s+/g, '').length;
+        const contentFactor = len >= SHORT_THRESHOLD
+            ? 1.0
+            : Math.max(0.3, Math.sqrt(len / SHORT_THRESHOLD));
+        raw.push(posWeight * contentFactor);
+    }
+
+    const sum = raw.reduce((a, b) => a + b, 0) || 1;
+    return raw.map(w => w / sum);
 }
 
 // ═══════════════════════════════════════════════════════════════════════════
@@ -247,19 +260,16 @@ function buildQuerySegments(chat, count, excludeLastAi, pendingUserMessage = nul
         }
     }
 
-    return messages.slice(-count).map((m, idx, arr) => {
+    return messages.slice(-count)
-        const speaker = m.name || (m.is_user ? (name1 || "用户") : "角色");
+        .map((m) => cleanForRecall(m.mes) || '')
-        const clean = cleanForRecall(m.mes);
+        .filter(Boolean);
-        if (!clean) return '';
-        const limit = idx === arr.length - 1 ? CONFIG.QUERY_MAX_CHARS : CONFIG.QUERY_CONTEXT_CHARS;
-        return `${speaker}: ${clean.slice(0, limit)}`;
-    }).filter(Boolean);
 }
 
 async function embedWeightedQuery(segments, vectorConfig) {
     if (!segments?.length) return null;
 
-    const weights = buildExpDecayWeights(segments.length, CONFIG.QUERY_DECAY_BETA);
+    const weights = buildContentAwareWeights(segments, CONFIG.QUERY_DECAY_BETA);
+
     const vecs = await embed(segments, vectorConfig);
     const dims = vecs?.[0]?.length || 0;
     if (!dims) return null;
@@ -377,19 +387,6 @@ function expandByFacts(presentEntities, facts, maxDepth = 2) {
 // 实体权重归一化（用于加分分配）
 // ═══════════════════════════════════════════════════════════════════════════
 
-function normalizeEntityWeights(queryEntityWeights) {
-    if (!queryEntityWeights?.size) return new Map();
-
-    const total = Array.from(queryEntityWeights.values()).reduce((a, b) => a + b, 0);
-    if (total <= 0) return new Map();
-
-    const normalized = new Map();
-    for (const [entity, weight] of queryEntityWeights) {
-        normalized.set(entity, weight / total);
-    }
-    return normalized;
-}
-
 // ═══════════════════════════════════════════════════════════════════════════
 // 文本路 Query 构建（分层高信号词）
 // ═══════════════════════════════════════════════════════════════════════════
@@ -548,7 +545,167 @@ function mmrSelect(candidates, k, lambda, getVector, getScore) {
 // L1 Chunks 检索
 // ═══════════════════════════════════════════════════════════════════════════
 
-async function searchChunks(queryVector, vectorConfig, l0FloorBonus = new Map(), lastSummarizedFloor = -1) {
+async function searchEvents(queryVector, queryTextForSearch, allEvents, vectorConfig, store, queryEntitySet, l0FloorBonus = new Map()) {
+    const { chatId } = getContext();
+    if (!chatId || !queryVector?.length) return [];
+
+    const meta = await getMeta(chatId);
+    const fp = getEngineFingerprint(vectorConfig);
+    if (meta.fingerprint && meta.fingerprint !== fp) return [];
+
+    const eventVectors = await getAllEventVectors(chatId);
+    const vectorMap = new Map(eventVectors.map(v => [v.eventId, v.vector]));
+    if (!vectorMap.size) return [];
+
+    // 构建/更新文本索引
+    const revision = `${chatId}:${store?.updatedAt || 0}:${allEvents.length}`;
+    ensureEventTextIndex(allEvents, revision);
+
+    // 文本路检索
+    const textRanked = searchEventsByText(queryTextForSearch, CONFIG.TEXT_SEARCH_LIMIT);
+    const textGapInfo = textRanked._gapInfo || null;
+
+    // 向量路检索
+    const scored = (allEvents || []).map((event, idx) => {
+        const v = vectorMap.get(event.id);
+        const rawSim = v ? cosineSimilarity(queryVector, v) : 0;
+
+        let bonus = 0;
+
+        // L0 加权
+        const range = parseFloorRange(event.summary);
+        if (range) {
+            for (let f = range.start; f <= range.end; f++) {
+                if (l0FloorBonus.has(f)) {
+                    bonus += l0FloorBonus.get(f);
+                    break;
+                }
+            }
+        }
+
+        const participants = (event.participants || []).map(p => normalize(p));
+        const hasPresent = participants.some(p => queryEntitySet.has(p));
+
+        return {
+            _id: event.id,
+            _idx: idx,
+            event,
+            rawSim,
+            finalScore: rawSim + bonus,
+            vector: v,
+            _hasPresent: hasPresent,
+        };
+    });
+
+    const rawSimById = new Map(scored.map(s => [s._id, s.rawSim]));
+    const hasPresentById = new Map(scored.map(s => [s._id, s._hasPresent]));
+
+    const preFilterDistribution = {
+        total: scored.length,
+        '0.85+': scored.filter(s => s.finalScore >= 0.85).length,
+        '0.7-0.85': scored.filter(s => s.finalScore >= 0.7 && s.finalScore < 0.85).length,
+        '0.6-0.7': scored.filter(s => s.finalScore >= 0.6 && s.finalScore < 0.7).length,
+        '0.5-0.6': scored.filter(s => s.finalScore >= 0.5 && s.finalScore < 0.6).length,
+        '<0.5': scored.filter(s => s.finalScore < 0.5).length,
+        passThreshold: scored.filter(s => s.finalScore >= CONFIG.MIN_SIMILARITY_EVENT).length,
+        threshold: CONFIG.MIN_SIMILARITY_EVENT,
+    };
+
+    const candidates = scored
+        .filter(s => s.finalScore >= CONFIG.MIN_SIMILARITY_EVENT)
+        .sort((a, b) => b.finalScore - a.finalScore)
+        .slice(0, CONFIG.CANDIDATE_EVENTS);
+
+    const vectorRanked = candidates.map(s => ({
+        event: s.event,
+        similarity: s.finalScore,
+        rawSim: s.rawSim,
+        vector: s.vector,
+    }));
+
+    const eventById = new Map(allEvents.map(e => [e.id, e]));
+    const fused = fuseEventsByRRF(vectorRanked, textRanked, eventById);
+
+    // TEXT-only 质量门槛
+    const textOnlyStats = {
+        total: 0,
+        passedSoftCheck: 0,
+        filtered: 0,
+    };
+
+    const filtered = fused.filter(x => {
+        if (x.type !== 'TEXT') return true;
+
+        textOnlyStats.total++;
+
+        const sim = x.rawSim || rawSimById.get(x.id) || 0;
+        if (sim >= CONFIG.TEXT_SOFT_MIN_SIM) {
+            textOnlyStats.passedSoftCheck++;
+            return true;
+        }
+
+        textOnlyStats.filtered++;
+        return false;
+    });
+
+    const mmrInput = filtered.slice(0, CONFIG.CANDIDATE_EVENTS).map(x => ({
+        ...x,
+        _id: x.id,
+    }));
+
+    const mmrOutput = mmrSelect(
+        mmrInput,
+        CONFIG.MAX_EVENTS,
+        CONFIG.MMR_LAMBDA,
+        c => c.vector || null,
+        c => c.rrf
+    );
+
+    // TEXT-only 限额（MMR 后执行）
+    let textOnlyCount = 0;
+    let textOnlyTruncated = 0;
+
+    const finalResults = mmrOutput.filter(x => {
+        if (x.type !== 'TEXT') return true;
+
+        if (textOnlyCount < CONFIG.TEXT_TOTAL_MAX) {
+            textOnlyCount++;
+            return true;
+        }
+
+        textOnlyTruncated++;
+        return false;
+    });
+
+    textOnlyStats.finalIncluded = textOnlyCount;
+    textOnlyStats.truncatedByLimit = textOnlyTruncated;
+
+    const results = finalResults.map(x => ({
+        event: x.event,
+        similarity: x.rrf,
+        _recallType: hasPresentById.get(x.event?.id) ? 'DIRECT' : 'SIMILAR',
+        _recallReason: x.type,
+        _rrfDetail: { vRank: x.vRank, tRank: x.tRank, rrf: x.rrf },
+        _rawSim: rawSimById.get(x.event?.id) || 0,
+    }));
+
+    if (results.length > 0) {
+        results[0]._preFilterDistribution = preFilterDistribution;
+        results[0]._rrfStats = {
+            vectorCount: vectorRanked.length,
+            textCount: textRanked.length,
+            hybridCount: fused.filter(x => x.type === 'HYBRID').length,
+            vectorOnlyCount: fused.filter(x => x.type === 'VECTOR').length,
+            textOnlyTotal: textOnlyStats.total,
+        };
+        results[0]._textOnlyStats = textOnlyStats;
+        results[0]._textGapInfo = textGapInfo;
+    }
+
+    return results;
+}
+
+async function searchChunks(queryVector, vectorConfig, l0FloorBonus = new Map(), lastSummarizedFloor = -1, textSearchParams = null) {
     const { chatId } = getContext();
     if (!chatId || !queryVector?.length) return [];
 
@@ -577,6 +734,58 @@ async function searchChunks(queryVector, vectorConfig, l0FloorBonus = new Map(),
         };
     });
 
+    // 文本路补充（仅待整理区）
+    let textL1Stats = null;
+    const store = getSummaryStore();
+    const keepVisible = store?.keepVisibleCount ?? 3;
+    const recentStart = lastSummarizedFloor + 1;
+    const recentEnd = (meta?.lastChunkFloor ?? -1) - keepVisible;
+
+    if (textSearchParams && recentEnd >= recentStart && recentEnd >= 0) {
+        const { queryEntities, rareTerms } = textSearchParams;
+        const textQuery = [...(queryEntities || []), ...(rareTerms || [])].join(' ');
+
+        if (textQuery.trim()) {
+            const allChunks = await getAllChunks(chatId);
+            const recentChunks = allChunks.filter(c => c.floor >= recentStart && c.floor <= recentEnd);
+
+            if (recentChunks.length > 0) {
+                const revision = `${chatId}:chunk:${recentEnd}`;
+                ensureChunkTextIndex(recentChunks, revision);
+
+                const textHits = searchChunksByText(textQuery, recentStart, recentEnd, 20);
+
+                textL1Stats = {
+                    range: `${recentStart + 1}~${recentEnd + 1}`,
+                    candidates: recentChunks.length,
+                    hits: textHits.length,
+                };
+
+                for (const hit of textHits) {
+                    const existingIdx = scored.findIndex(s => s.chunkId === hit.chunkId);
+
+                    if (existingIdx >= 0) {
+                        scored[existingIdx]._hasTextHit = true;
+                        scored[existingIdx]._textRank = hit.textRank;
+                    } else {
+                        scored.push({
+                            _id: hit.chunkId,
+                            chunkId: hit.chunkId,
+                            floor: hit.floor,
+                            chunkIdx: 0,
+                            similarity: CONFIG.MIN_SIMILARITY_CHUNK_RECENT,
+                            _baseSimilarity: 0,
+                            _l0Bonus: 0,
+                            _recallReason: 'TEXT_L1',
+                            _textRank: hit.textRank,
+                            vector: null,
+                        });
+                    }
+                }
+            }
+        }
+    }
+
     const candidates = scored
         .filter(s => {
             const threshold = s.floor > lastSummarizedFloor
@@ -599,6 +808,7 @@ async function searchChunks(queryVector, vectorConfig, l0FloorBonus = new Map(),
             '0.55-0.6': scored.filter(s => s.similarity >= 0.55 && s.similarity < 0.6).length,
             '<0.55': scored.filter(s => s.similarity < 0.55).length,
         },
+        textL1: textL1Stats,
     };
 
     const dynamicK = Math.min(CONFIG.MAX_CHUNKS, candidates.length);
@@ -636,6 +846,8 @@ async function searchChunks(queryVector, vectorConfig, l0FloorBonus = new Map(),
             isUser: chunk.isUser,
             text: chunk.text,
             similarity: item.similarity,
+            _recallReason: item._recallReason,
+            _textRank: item._textRank,
         };
     }).filter(Boolean);
 
@@ -646,184 +858,6 @@ async function searchChunks(queryVector, vectorConfig, l0FloorBonus = new Map(),
     return results;
 }
 
-// ═══════════════════════════════════════════════════════════════════════════
-// L2 Events 检索（RRF 混合 + MMR 后置）
-// ═══════════════════════════════════════════════════════════════════════════
-
-async function searchEvents(queryVector, queryTextForSearch, allEvents, vectorConfig, store, normalizedEntityWeights, l0FloorBonus = new Map()) {
-    const { chatId } = getContext();
-    if (!chatId || !queryVector?.length) return [];
-
-    const meta = await getMeta(chatId);
-    const fp = getEngineFingerprint(vectorConfig);
-    if (meta.fingerprint && meta.fingerprint !== fp) return [];
-
-    const eventVectors = await getAllEventVectors(chatId);
-    const vectorMap = new Map(eventVectors.map(v => [v.eventId, v.vector]));
-    if (!vectorMap.size) return [];
-
-    // 构建/更新文本索引
-    const revision = `${chatId}:${store?.updatedAt || 0}:${allEvents.length}`;
-    ensureEventTextIndex(allEvents, revision);
-
-    // 文本路检索
-    const textRanked = searchEventsByText(queryTextForSearch, CONFIG.TEXT_SEARCH_LIMIT);
-    const textGapInfo = textRanked._gapInfo || null;
-
-    // ═══════════════════════════════════════════════════════════════════════
-    // 向量路检索（只保留 L0 加权）
-    // ═══════════════════════════════════════════════════════════════════════
-
-    const ENTITY_BONUS_POOL = 0.10;
-
-    const scored = (allEvents || []).map((event, idx) => {
-        const v = vectorMap.get(event.id);
-        const rawSim = v ? cosineSimilarity(queryVector, v) : 0;
-
-        let bonus = 0;
-
-        // L0 加权
-        const range = parseFloorRange(event.summary);
-        if (range) {
-            for (let f = range.start; f <= range.end; f++) {
-                if (l0FloorBonus.has(f)) {
-                    bonus += l0FloorBonus.get(f);
-                    break;
-                }
-            }
-        }
-
-        const participants = (event.participants || []).map(p => normalize(p));
-        let maxEntityWeight = 0;
-        for (const p of participants) {
-            const w = normalizedEntityWeights.get(p) || 0;
-            if (w > maxEntityWeight) {
-                maxEntityWeight = w;
-            }
-        }
-        const entityBonus = ENTITY_BONUS_POOL * maxEntityWeight;
-        bonus += entityBonus;
-
-        return {
-            _id: event.id,
-            _idx: idx,
-            event,
-            rawSim,
-            finalScore: rawSim + bonus,
-            vector: v,
-            _entityBonus: entityBonus,
-            _hasPresent: maxEntityWeight > 0,
-        };
-    });
-
-    const rawSimById = new Map(scored.map(s => [s._id, s.rawSim]));
-    const entityBonusById = new Map(scored.map(s => [s._id, s._entityBonus]));
-    const hasPresentById = new Map(scored.map(s => [s._id, s._hasPresent]));
-
-    const preFilterDistribution = {
-        total: scored.length,
-        '0.85+': scored.filter(s => s.finalScore >= 0.85).length,
-        '0.7-0.85': scored.filter(s => s.finalScore >= 0.7 && s.finalScore < 0.85).length,
-        '0.6-0.7': scored.filter(s => s.finalScore >= 0.6 && s.finalScore < 0.7).length,
-        '0.5-0.6': scored.filter(s => s.finalScore >= 0.5 && s.finalScore < 0.6).length,
-        '<0.5': scored.filter(s => s.finalScore < 0.5).length,
-        passThreshold: scored.filter(s => s.finalScore >= CONFIG.MIN_SIMILARITY_EVENT).length,
-        threshold: CONFIG.MIN_SIMILARITY_EVENT,
-    };
-
-    const candidates = scored
-        .filter(s => s.finalScore >= CONFIG.MIN_SIMILARITY_EVENT)
-        .sort((a, b) => b.finalScore - a.finalScore)
-        .slice(0, CONFIG.CANDIDATE_EVENTS);
-
-    const vectorRanked = candidates.map(s => ({
-        event: s.event,
-        similarity: s.finalScore,
-        rawSim: s.rawSim,
-        vector: s.vector,
-    }));
-
-    const eventById = new Map(allEvents.map(e => [e.id, e]));
-    const fused = fuseEventsByRRF(vectorRanked, textRanked, eventById);
-
-    const textOnlyStats = {
-        total: 0,
-        passedSoftCheck: 0,
-        filtered: 0,
-        finalIncluded: 0,
-        truncatedByLimit: 0,
-    };
-
-    const filtered = fused.filter(x => {
-        if (x.type !== 'TEXT') return true;
-
-        textOnlyStats.total++;
-        const sim = x.rawSim || rawSimById.get(x.id) || 0;
-        if (sim >= CONFIG.TEXT_SOFT_MIN_SIM) {
-            textOnlyStats.passedSoftCheck++;
-            return true;
-        }
-
-        textOnlyStats.filtered++;
-        return false;
-    });
-
-    const mmrInput = filtered.slice(0, CONFIG.CANDIDATE_EVENTS).map(x => ({
-        ...x,
-        _id: x.id,
-    }));
-
-    const mmrOutput = mmrSelect(
-        mmrInput,
-        CONFIG.MAX_EVENTS,
-        CONFIG.MMR_LAMBDA,
-        c => c.vector || null,
-        c => c.rrf
-    );
-
-    let textOnlyCount = 0;
-    const finalResults = mmrOutput.filter(x => {
-        if (x.type !== 'TEXT') return true;
-        if (textOnlyCount < CONFIG.TEXT_TOTAL_MAX) {
-            textOnlyCount++;
-            return true;
-        }
-        textOnlyStats.truncatedByLimit++;
-        return false;
-    });
-    textOnlyStats.finalIncluded = textOnlyCount;
-
-    const results = finalResults.map(x => ({
-        event: x.event,
-        similarity: x.rrf,
-        _recallType: hasPresentById.get(x.event?.id) ? 'DIRECT' : 'SIMILAR',
-        _recallReason: x.type,
-        _rrfDetail: { vRank: x.vRank, tRank: x.tRank, rrf: x.rrf },
-        _entityBonus: entityBonusById.get(x.event?.id) || 0,
-        _rawSim: rawSimById.get(x.event?.id) || 0,
-    }));
-
-    // 统计信息附加到第一条结果
-    if (results.length > 0) {
-        results[0]._preFilterDistribution = preFilterDistribution;
-        results[0]._rrfStats = {
-            vectorCount: vectorRanked.length,
-            textCount: textRanked.length,
-            hybridCount: fused.filter(x => x.type === 'HYBRID').length,
-            vectorOnlyCount: fused.filter(x => x.type === 'VECTOR').length,
-            textOnlyTotal: textOnlyStats.total,
-        };
-        results[0]._textOnlyStats = textOnlyStats;
-        results[0]._textGapInfo = textGapInfo;
-    }
-
-    return results;
-}
-
-// ═══════════════════════════════════════════════════════════════════════════
-// 日志
-// ═══════════════════════════════════════════════════════════════════════════
-
 function formatRecallLog({
     elapsed,
     segments,
@@ -831,7 +865,7 @@ function formatRecallLog({
     chunkResults,
     eventResults,
     allEvents,
-    normalizedEntityWeights = new Map(),
+    queryEntities = [],
     causalEvents = [],
     chunkPreFilterStats = null,
     l0Results = [],
@@ -840,15 +874,15 @@ function formatRecallLog({
     textQueryBreakdown = null,
 }) {
     const lines = [
-        '\u2554' + '\u2550'.repeat(62) + '\u2557',
+        '╔' + '═'.repeat(62) + '╗',
-        '\u2551                    记忆召回报告                              \u2551',
+        '║                    记忆召回报告                              ║',
-        '\u2560' + '\u2550'.repeat(62) + '\u2563',
+        '╠' + '═'.repeat(62) + '╣',
-        `\u2551  耗时: ${elapsed}ms`,
+        `║  耗时: ${elapsed}ms`,
-        '\u255a' + '\u2550'.repeat(62) + '\u255d',
+        '╚' + '═'.repeat(62) + '╝',
         '',
-        '\u250c' + '\u2500'.repeat(61) + '\u2510',
+        '┌' + '─'.repeat(61) + '┐',
-        '\u2502 【查询构建】最近 5 条消息，指数衰减加权 (β=0.7)              \u2502',
+        `│ 【查询构建】最近 ${CONFIG.QUERY_MSG_COUNT} 条，内容感知加权 (β=${CONFIG.QUERY_DECAY_BETA}) │`,
-        '\u2514' + '\u2500'.repeat(61) + '\u2518',
+        '└' + '─'.repeat(61) + '┘',
     ];
 
     const segmentsSorted = segments.map((s, i) => ({
@@ -858,25 +892,19 @@ function formatRecallLog({
     })).sort((a, b) => b.weight - a.weight);
 
     segmentsSorted.forEach((s, rank) => {
-        const bar = '\u2588'.repeat(Math.round(s.weight * 20));
+        const bar = '█'.repeat(Math.round(s.weight * 20));
         const preview = s.text.length > 60 ? s.text.slice(0, 60) + '...' : s.text;
         const marker = rank === 0 ? ' ◀ 主导' : '';
         lines.push(`  ${(s.weight * 100).toFixed(1).padStart(5)}% ${bar.padEnd(12)} ${preview}${marker}`);
     });
 
     lines.push('');
-    lines.push('\u250c' + '\u2500'.repeat(61) + '\u2510');
+    lines.push('┌' + '─'.repeat(61) + '┐');
-    lines.push('\u2502 【提取实体】                                                 \u2502');
+    lines.push('│ 【提取实体】                                                 │');
-    lines.push('\u2514' + '\u2500'.repeat(61) + '\u2518');
+    lines.push('└' + '─'.repeat(61) + '┘');
 
-    if (normalizedEntityWeights?.size) {
+    if (queryEntities?.length) {
-        const sorted = Array.from(normalizedEntityWeights.entries())
+        lines.push(`  焦点: ${queryEntities.slice(0, 8).join('、')}${queryEntities.length > 8 ? ' ...' : ''}`);
-            .sort((a, b) => b[1] - a[1])
-            .slice(0, 8);
-        const formatted = sorted
-            .map(([e, w]) => `${e}(${(w * 100).toFixed(0)}%)`)
-            .join(' | ');
-        lines.push(`  ${formatted}`);
     } else {
         lines.push('  (无)');
     }
@@ -885,9 +913,9 @@ function formatRecallLog({
     }
 
     lines.push('');
-    lines.push('\u250c' + '\u2500'.repeat(61) + '\u2510');
+    lines.push('┌' + '─'.repeat(61) + '┐');
-    lines.push('\u2502 【文本路 Query 构成】                                        \u2502');
+    lines.push('│ 【文本路 Query 构成】                                        │');
-    lines.push('\u2514' + '\u2500'.repeat(61) + '\u2518');
+    lines.push('└' + '─'.repeat(61) + '┘');
 
     if (textQueryBreakdown) {
         const bd = textQueryBreakdown;
@@ -919,23 +947,9 @@ function formatRecallLog({
     }
 
     lines.push('');
-    lines.push('  实体归一化（用于加分）:');
+    lines.push('┌' + '─'.repeat(61) + '┐');
-    if (normalizedEntityWeights?.size) {
+    lines.push('│ 【召回统计】                                                 │');
-        const sorted = Array.from(normalizedEntityWeights.entries())
+    lines.push('└' + '─'.repeat(61) + '┘');
-            .sort((a, b) => b[1] - a[1])
-            .slice(0, 8);
-        const formatted = sorted
-            .map(([e, w]) => `${e}(${(w * 100).toFixed(0)}%)`)
-            .join(' | ');
-        lines.push(`    ${formatted}`);
-    } else {
-        lines.push('    (无)');
-    }
-
-    lines.push('');
-    lines.push('\u250c' + '\u2500'.repeat(61) + '\u2510');
-    lines.push('\u2502 【召回统计】                                                 \u2502');
-    lines.push('\u2514' + '\u2500'.repeat(61) + '\u2518');
 
     // L0
     const l0Floors = [...new Set(l0Results.map(r => r.floor))].sort((a, b) => a - b);
@@ -953,6 +967,11 @@ function formatRecallLog({
         const dist = chunkPreFilterStats.distribution || {};
         lines.push(`    全量: ${chunkPreFilterStats.total} 条 | 通过阈值(远期≥${chunkPreFilterStats.thresholdRemote}, 待整理≥${chunkPreFilterStats.thresholdRecent}): ${chunkPreFilterStats.passThreshold} 条 | 最终: ${chunkResults.length} 条`);
         lines.push(`    匹配度: 0.8+: ${dist['0.8+'] || 0} | 0.7-0.8: ${dist['0.7-0.8'] || 0} | 0.6-0.7: ${dist['0.6-0.7'] || 0}`);
+
+        const textL1 = chunkPreFilterStats.textL1;
+        if (textL1) {
+            lines.push(`    文本路补充（待整理区）: 范围 ${textL1.range}楼 | 候选 ${textL1.candidates} 条 | 命中 ${textL1.hits} 条`);
+        }
     } else {
         lines.push(`    选入: ${chunkResults.length} 条`);
     }
@@ -988,9 +1007,6 @@ function formatRecallLog({
             lines.push(`    ${i + 1}. [${id}] ${title.padEnd(25)} sim=${sim} tRank=${tRank}`);
         });
     }
-    const entityBoostedEvents = eventResults.filter(e => e._entityBonus > 0).length;
-    lines.push('');
-    lines.push(`    实体加分事件: ${entityBoostedEvents} 条`);
 
     if (textGapInfo) {
         lines.push('');
@@ -1002,7 +1018,6 @@ function formatRecallLog({
         }
     }
 
-    // Causal
     if (causalEvents.length) {
         const maxRefs = Math.max(...causalEvents.map(c => c.chainFrom?.length || 0));
         const maxDepth = Math.max(...causalEvents.map(c => c.depth || 0));
@@ -1012,13 +1027,8 @@ function formatRecallLog({
     }
 
     lines.push('');
-    return lines.join('\n');
+    return lines.join("\n");
 }
-
-// ═══════════════════════════════════════════════════════════════════════════
-// 主入口
-// ═══════════════════════════════════════════════════════════════════════════
-
 export async function recallMemory(queryText, allEvents, vectorConfig, options = {}) {
     const T0 = performance.now();
     const { chat } = getContext();
@@ -1049,9 +1059,9 @@ export async function recallMemory(queryText, allEvents, vectorConfig, options =
     const lexicon = buildEntityLexicon(store, allEvents);
     const queryEntityWeights = extractEntitiesWithWeights(segments, weights, lexicon);
     const queryEntities = Array.from(queryEntityWeights.keys());
+    const queryEntitySet = new Set(queryEntities.map(normalize));
     const facts = getFacts(store);
     const expandedTerms = expandByFacts(queryEntities, facts, 2);
-    const normalizedEntityWeights = normalizeEntityWeights(queryEntityWeights);
 
     let queryTextForSearch = '';
     let textQueryBreakdown = null;
@@ -1079,8 +1089,11 @@ export async function recallMemory(queryText, allEvents, vectorConfig, options =
     }
 
     const [chunkResults, eventResults] = await Promise.all([
-        searchChunks(queryVector, vectorConfig, l0FloorBonus, lastSummarizedFloor),
+        searchChunks(queryVector, vectorConfig, l0FloorBonus, lastSummarizedFloor, {
-        searchEvents(queryVector, queryTextForSearch, allEvents, vectorConfig, store, normalizedEntityWeights, l0FloorBonus),
+            queryEntities,
+            rareTerms: textQueryBreakdown?.rareTerms || [],
+        }),
+        searchEvents(queryVector, queryTextForSearch, allEvents, vectorConfig, store, queryEntitySet, l0FloorBonus),
     ]);
 
     const chunkPreFilterStats = chunkResults._preFilterStats || null;
@@ -1118,7 +1131,7 @@ export async function recallMemory(queryText, allEvents, vectorConfig, options =
         chunkResults: mergedChunks,
         eventResults,
         allEvents,
-        normalizedEntityWeights,
+        queryEntities,
         causalEvents: causalEventsTruncated,
         chunkPreFilterStats,
         l0Results,
@@ -1149,3 +1162,8 @@ export function buildQueryText(chat, count = 2, excludeLastAi = false) {
         return `${speaker}: ${text.slice(0, 500)}`;
     }).filter(Boolean).join('\n');
 }
+
+
+
+
+

modules/story-summary/vector/retrieval/text-search.js

@@ -1,6 +1,6 @@
 // text-search.js - 最终版
 
-import MiniSearch from '../../../libs/minisearch.mjs';
+import MiniSearch from '../../../../libs/minisearch.mjs';
 
 const STOP_WORDS = new Set([
     '的', '了', '是', '在', '和', '与', '或', '但', '而', '却',
@@ -106,7 +106,7 @@ export function ensureEventTextIndex(events, revision) {
  *
  * 参考：帕累托法则（80/20 法则）在信息检索中的应用
  */
-function dynamicTopK(scores, coverage = 0.90, minK = 15, maxK = 80) {
+export function dynamicTopK(scores, coverage = 0.90, minK = 15, maxK = 80) {
     if (!scores.length) return 0;
 
     const total = scores.reduce((a, b) => a + b, 0);
@@ -171,3 +171,67 @@ export function clearEventTextIndex() {
     idx = null;
     lastRevision = null;
 }
+
+// ---------------------------------------------------------------------------
+// Chunk 文本索引（待整理区 L1 补充）
+// ---------------------------------------------------------------------------
+
+let chunkIdx = null;
+let chunkIdxRevision = null;
+
+export function ensureChunkTextIndex(chunks, revision) {
+    if (chunkIdx && revision === chunkIdxRevision) return;
+
+    try {
+        chunkIdx = new MiniSearch({
+            fields: ['text'],
+            storeFields: ['chunkId', 'floor'],
+            tokenize,
+            searchOptions: { tokenize },
+        });
+
+        chunkIdx.addAll(chunks.map(c => ({
+            id: c.chunkId,
+            chunkId: c.chunkId,
+            floor: c.floor,
+            text: c.text || '',
+        })));
+
+        chunkIdxRevision = revision;
+    } catch (e) {
+        console.error('[text-search] Chunk index build failed:', e);
+        chunkIdx = null;
+    }
+}
+
+export function searchChunksByText(query, floorMin, floorMax, limit = 20) {
+    if (!chunkIdx || !query?.trim()) return [];
+
+    try {
+        const results = chunkIdx.search(query, {
+            fuzzy: false,
+            prefix: false,
+        });
+
+        const filtered = results.filter(r => r.floor >= floorMin && r.floor <= floorMax);
+        if (!filtered.length) return [];
+
+        const scores = filtered.map(r => r.score);
+        const k = dynamicTopK(scores, 0.85, 5, limit);
+
+        return filtered.slice(0, k).map((r, i) => ({
+            chunkId: r.chunkId,
+            floor: r.floor,
+            textRank: i + 1,
+            score: r.score,
+        }));
+    } catch (e) {
+        console.error('[text-search] Chunk search failed:', e);
+        return [];
+    }
+}
+
+export function clearChunkTextIndex() {
+    chunkIdx = null;
+    chunkIdxRevision = null;
+}

modules/story-summary/vector/retrieval/tokenizer.js

@@ -1,5 +1,5 @@
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
-import { extensionFolderPath } from '../../../core/constants.js';
+import { extensionFolderPath } from '../../../../core/constants.js';
 
 const MODULE_ID = 'tokenizer';
 

modules/story-summary/vector/storage/chunk-store.js

@@ -8,7 +8,7 @@ import {
     chunkVectorsTable,
     eventVectorsTable,
     CHUNK_MAX_TOKENS,
-} from '../data/db.js';
+} from '../../data/db.js';
 
 // ═══════════════════════════════════════════════════════════════════════════
 // 工具函数

modules/story-summary/vector/storage/state-store.js

@@ -4,11 +4,11 @@
 // StateVector 存 IndexedDB（可重建）
 // ═══════════════════════════════════════════════════════════════════════════
 
-import { saveMetadataDebounced } from '../../../../../../extensions.js';
+import { saveMetadataDebounced } from '../../../../../../../extensions.js';
-import { chat_metadata } from '../../../../../../../script.js';
+import { chat_metadata } from '../../../../../../../../script.js';
-import { stateVectorsTable } from '../data/db.js';
+import { stateVectorsTable } from '../../data/db.js';
-import { EXT_ID } from '../../../core/constants.js';
+import { EXT_ID } from '../../../../core/constants.js';
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
 
 const MODULE_ID = 'state-store';
 

modules/story-summary/vector/storage/vector-io.js

@@ -3,9 +3,9 @@
 // 向量数据导入导出（当前 chatId 级别）
 // ═══════════════════════════════════════════════════════════════════════════
 
-import { zipSync, unzipSync, strToU8, strFromU8 } from '../../../libs/fflate.mjs';
+import { zipSync, unzipSync, strToU8, strFromU8 } from '../../../../libs/fflate.mjs';
-import { getContext } from '../../../../../../extensions.js';
+import { getContext } from '../../../../../../../extensions.js';
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
 import {
     getMeta,
     updateMeta,
@@ -26,8 +26,8 @@ import {
     saveStateVectors,
     clearStateVectors,
 } from './state-store.js';
-import { getEngineFingerprint } from './embedder.js';
+import { getEngineFingerprint } from '../utils/embedder.js';
-import { getVectorConfig } from '../data/config.js';
+import { getVectorConfig } from '../../data/config.js';
 
 const MODULE_ID = 'vector-io';
 const EXPORT_VERSION = 1;

modules/story-summary/vector/utils/embedder.js

@@ -3,7 +3,7 @@
 // 统一的向量生成接口（本地模型 / 在线服务）
 // ═══════════════════════════════════════════════════════════════════════════
 
-import { xbLog } from '../../../core/debug-core.js';
+import { xbLog } from '../../../../core/debug-core.js';
 
 const MODULE_ID = 'embedding';
 

modules/story-summary/vector/utils/text-filter.js

@@ -3,7 +3,7 @@
 // 跳过用户定义的「起始→结束」区间
 // ═══════════════════════════════════════════════════════════════════════════
 
-import { getTextFilterRules } from '../data/config.js';
+import { getTextFilterRules } from '../../data/config.js';
 
 /**
  * 转义正则特殊字符