X/LittleWhiteBox
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update recall metrics and context pairing

Browse Source
This commit is contained in:
bielie
2026-02-10 00:18:51 +08:00
parent da1e3088eb
commit 3af76a9651
3 changed files with 657 additions and 750 deletions
Download Patch File Download Diff File Expand all files Collapse all files

743
modules/story-summary/generate/prompt.js
View File

@@ -1,10 +1,18 @@
// ═══════════════════════════════════════════════════════════════════════════
// Story Summary - Prompt Injection (v4 - 统一命名)
// Story Summary - Prompt Injection (v5 - Two-Stage: L0 Locate → L1 Evidence)
//
// 命名规范：
// - 存储层用 L0/L1/L2/L3（StateAtom/Chunk/Event/Fact）
// - 装配层用语义名称：constraint/event/evidence/arc
//
// 架构变更（v4 → v5）：
// - L0 和 L1 不再在同一个池子竞争
// - recall.js 返回 {l0Selected[], l1ByFloor: Map} 而非 evidenceChunks[]
// - 装配层按 L2→L0→L1 层级组织
// - 预算以"L0 + USER top-1 + AI top-1"为原子单元
// - 孤立 L1（无对应 L0）丢弃
// - 孤立 L0（无对应 L1）保留
//
// 职责：
// - 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
@@ -15,7 +23,7 @@ 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 } from "../vector/retrieval/recall.js";
import { getChunksByFloors, getAllChunkVectors, getAllEventVectors, getMeta } from "../vector/storage/chunk-store.js";
import { getMeta } from "../vector/storage/chunk-store.js";
// Metrics
import { formatMetricsLog, detectIssues } from "../vector/retrieval/metrics.js";
@@ -79,23 +87,6 @@ function pushWithBudget(lines, text, state) {
return true;
}
/**
* 计算余弦相似度
* @param {number[]} a - 向量A
* @param {number[]} b - 向量B
* @returns {number} 相似度
*/
function cosineSimilarity(a, b) {
if (!a?.length || !b?.length || a.length !== b.length) return 0;
let dot = 0, nA = 0, nB = 0;
for (let i = 0; i < a.length; i++) {
dot += a[i] * b[i];
nA += a[i] * a[i];
nB += b[i] * b[i];
}
return nA && nB ? dot / (Math.sqrt(nA) * Math.sqrt(nB)) : 0;
}
/**
* 解析事件摘要中的楼层范围
* @param {string} summary - 事件摘要
@@ -134,46 +125,27 @@ function normalize(s) {
.toLowerCase();
}
// ─────────────────────────────────────────────────────────────────────────────
// 上下文配对工具函数
// ─────────────────────────────────────────────────────────────────────────────
/**
* 获取 chunk 的上下文楼层
* @param {object} chunk - chunk 对象
* @returns {number} 上下文楼层（-1 表示无）
* 获取事件排序键
* @param {object} event - 事件对象
* @returns {number} 排序键
*/
function getContextFloor(chunk) {
if (chunk.isAnchorVirtual) return -1;
return chunk.isUser ? chunk.floor + 1 : chunk.floor - 1;
function getEventSortKey(event) {
const r = parseFloorRange(event?.summary);
if (r) return r.start;
const m = String(event?.id || "").match(/evt-(\d+)/);
return m ? parseInt(m[1], 10) : Number.MAX_SAFE_INTEGER;
}
/**
* 选择上下文 chunk
* @param {object[]} candidates - 候选 chunks
* @param {object} mainChunk - 主 chunk
* @returns {object|null} 选中的上下文 chunk
* 重新编号事件文本
* @param {string} text - 原始文本
* @param {number} newIndex - 新编号
* @returns {string} 重新编号后的文本
*/
function pickContextChunk(candidates, mainChunk) {
if (!candidates?.length) return null;
const targetIsUser = !mainChunk.isUser;
const opposite = candidates.find(c => c.isUser === targetIsUser);
if (opposite) return opposite;
return candidates[0];
}
/**
* 格式化上下文 chunk 行
* @param {object} chunk - chunk 对象
* @param {boolean} isAbove - 是否在上方
* @returns {string} 格式化后的行
*/
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}`;
function renumberEventText(text, newIndex) {
const s = String(text || "");
return s.replace(/^(\s*)\d+(\.\s*(?:【)?)/, `$1${newIndex}$2`);
}
// ─────────────────────────────────────────────────────────────────────────────
@@ -259,10 +231,8 @@ function filterConstraintsByRelevance(facts, focusEntities, knownCharacters) {
const focusSet = new Set((focusEntities || []).map(normalize));
return facts.filter(f => {
// isState 的 facts 始终保留
if (f._isState === true) return true;
// 关系类 facts：检查 from/to 是否在焦点中
if (isRelationFact(f)) {
const from = normalize(f.s);
const target = parseRelationTarget(f.p);
@@ -272,7 +242,6 @@ function filterConstraintsByRelevance(facts, focusEntities, knownCharacters) {
return false;
}
// 其他 facts：检查主体是否在焦点中
const subjectNorm = normalize(f.s);
if (knownCharacters.has(subjectNorm)) {
return focusSet.has(subjectNorm);
@@ -326,28 +295,34 @@ function formatArcLine(arc) {
}
/**
* 格式化 evidence chunk 完整行
* @param {object} chunk - chunk 对象
* 格式化 L0 锚点行
* @param {object} l0 - L0 对象
* @returns {string} 格式化后的行
*/
function formatEvidenceFullLine(chunk) {
function formatL0Line(l0) {
return ` › #${l0.floor + 1} [📌] ${String(l0.text || l0.atom?.semantic || "").trim()}`;
}
/**
* 格式化 L1 chunk 行（挂在 L0 下方）
* @param {object} chunk - L1 chunk 对象
* @param {boolean} isContext - 是否为上下文（USER 侧）
* @returns {string} 格式化后的行
*/
function formatL1Line(chunk, isContext) {
const { name1, name2 } = getContext();
if (chunk.isAnchorVirtual) {
return `› #${chunk.floor + 1} [📌] ${String(chunk.text || "").trim()}`;
}
const speaker = chunk.isUser ? (name1 || "用户") : (chunk.speaker || name2 || "角色");
return `› #${chunk.floor + 1} [${speaker}] ${String(chunk.text || "").trim()}`;
const text = String(chunk.text || "").trim();
const symbol = isContext ? "┌" : "›";
return ` ${symbol} #${chunk.floor + 1} [${speaker}] ${text}`;
}
/**
* 格式化因果事件行
* @param {object} causalItem - 因果事件项
* @param {Map} causalById - 因果事件索引
* @returns {string} 格式化后的行
*/
function formatCausalEventLine(causalItem, causalById) {
function formatCausalEventLine(causalItem) {
const ev = causalItem?.event || {};
const depth = Math.max(1, Math.min(9, causalItem?._causalDepth || 1));
const indent = " │" + " ".repeat(depth - 1);
@@ -365,128 +340,128 @@ function formatCausalEventLine(causalItem, causalById) {
const body = `${summary}${floorHint ? ` ${floorHint}` : ""}`.trim();
lines.push(`${indent} ${body}`);
const evidence = causalItem._evidenceChunk;
if (evidence) {
const speaker = evidence.speaker || "角色";
const text = String(evidence.text || "").trim();
lines.push(`${indent} › #${evidence.floor + 1} [${speaker}] ${text}`);
}
return lines.join("\n");
}
/**
* 重新编号事件文本
* @param {string} text - 原始文本
* @param {number} newIndex - 新编号
* @returns {string} 重新编号后的文本
*/
function renumberEventText(text, newIndex) {
const s = String(text || "");
return s.replace(/^(\s*)\d+(\.\s*(?:【)?)/, `$1${newIndex}$2`);
}
/**
* 获取事件排序键
* @param {object} event - 事件对象
* @returns {number} 排序键
*/
function getEventSortKey(event) {
const r = parseFloorRange(event?.summary);
if (r) return r.start;
const m = String(event?.id || "").match(/evt-(\d+)/);
return m ? parseInt(m[1], 10) : Number.MAX_SAFE_INTEGER;
}
// ─────────────────────────────────────────────────────────────────────────────
// 按楼层分组装配 evidence（修复上下文重复）
// L0→L1 证据单元构建
// ─────────────────────────────────────────────────────────────────────────────
/**
* 按楼层装配 evidence
* @param {object[]} evidenceCandidates - 候选 evidence
* @param {Map} contextChunksByFloor - 上下文 chunks 索引
* @param {object} budget - 预算状态
* @returns {{lines: string[], anchorCount: number, contextPairsCount: number}}
* @typedef {object} EvidenceUnit
* @property {object} l0 - L0 锚点对象
* @property {object|null} userL1 - USER 侧 top-1 L1 chunk
* @property {object|null} aiL1 - AI 侧 top-1 L1 chunk
* @property {number} totalTokens - 整个单元的 token 估算
*/
function assembleEvidenceByFloor(evidenceCandidates, contextChunksByFloor, budget) {
if (!evidenceCandidates?.length) {
return { lines: [], anchorCount: 0, contextPairsCount: 0 };
}
// 1. 按楼层分组
const byFloor = new Map();
for (const c of evidenceCandidates) {
const arr = byFloor.get(c.floor) || [];
arr.push(c);
byFloor.set(c.floor, arr);
}
/**
* 为一个 L0 构建证据单元
* @param {object} l0 - L0 对象
* @param {Map<number, object>} l1ByFloor - 楼层→L1配对映射
* @returns {EvidenceUnit}
*/
function buildEvidenceUnit(l0, l1ByFloor) {
const pair = l1ByFloor.get(l0.floor);
const userL1 = pair?.userTop1 || null;
const aiL1 = pair?.aiTop1 || null;
// 2. 楼层内按 chunkIdx 排序
for (const [, chunks] of byFloor) {
chunks.sort((a, b) => (a.chunkIdx ?? 0) - (b.chunkIdx ?? 0));
}
// 预计算整个单元的 token 开销
let totalTokens = estimateTokens(formatL0Line(l0));
if (userL1) totalTokens += estimateTokens(formatL1Line(userL1, true));
if (aiL1) totalTokens += estimateTokens(formatL1Line(aiL1, false));
// 3. 按楼层顺序装配
const floorsSorted = Array.from(byFloor.keys()).sort((a, b) => a - b);
return { l0, userL1, aiL1, totalTokens };
}
/**
* 格式化一个证据单元为文本行
* @param {EvidenceUnit} unit - 证据单元
* @returns {string[]} 文本行数组
*/
function formatEvidenceUnit(unit) {
const lines = [];
let anchorCount = 0;
let contextPairsCount = 0;
lines.push(formatL0Line(unit.l0));
if (unit.userL1) {
lines.push(formatL1Line(unit.userL1, true));
}
if (unit.aiL1) {
lines.push(formatL1Line(unit.aiL1, false));
}
return lines;
}
for (const floor of floorsSorted) {
const chunks = byFloor.get(floor);
if (!chunks?.length) continue;
// ─────────────────────────────────────────────────────────────────────────────
// 事件证据收集
// ─────────────────────────────────────────────────────────────────────────────
// 分离锚点虚拟 chunks 和真实 chunks
const anchorChunks = chunks.filter(c => c.isAnchorVirtual);
const realChunks = chunks.filter(c => !c.isAnchorVirtual);
/**
* 为事件收集范围内的 L0 证据单元
* @param {object} eventObj - 事件对象
* @param {object[]} l0Selected - 所有选中的 L0
* @param {Map<number, object>} l1ByFloor - 楼层→L1配对映射
* @param {Set<string>} usedL0Ids - 已消费的 L0 ID 集合（会被修改）
* @returns {EvidenceUnit[]} 该事件的证据单元列表
*/
function collectEvidenceForEvent(eventObj, l0Selected, l1ByFloor, usedL0Ids) {
const range = parseFloorRange(eventObj?.summary);
if (!range) return [];
// 锚点直接输出（不需要上下文）
for (const c of anchorChunks) {
const line = formatEvidenceFullLine(c);
if (!pushWithBudget(lines, line, budget)) {
return { lines, anchorCount, contextPairsCount };
}
anchorCount++;
}
const units = [];
// 真实 chunks 按楼层统一处理
if (realChunks.length > 0) {
const firstChunk = realChunks[0];
const pairFloor = getContextFloor(firstChunk);
const pairCandidates = contextChunksByFloor.get(pairFloor) || [];
const contextChunk = pickContextChunk(pairCandidates, firstChunk);
for (const l0 of l0Selected) {
if (usedL0Ids.has(l0.id)) continue;
if (l0.floor < range.start || l0.floor > range.end) continue;
// 上下文在前
if (contextChunk && contextChunk.floor < floor) {
const contextLine = formatContextChunkLine(contextChunk, true);
if (!pushWithBudget(lines, contextLine, budget)) {
return { lines, anchorCount, contextPairsCount };
}
contextPairsCount++;
}
// 输出该楼层所有真实 chunks
for (const c of realChunks) {
const line = formatEvidenceFullLine(c);
if (!pushWithBudget(lines, line, budget)) {
return { lines, anchorCount, contextPairsCount };
}
}
// 上下文在后
if (contextChunk && contextChunk.floor > floor) {
const contextLine = formatContextChunkLine(contextChunk, false);
if (!pushWithBudget(lines, contextLine, budget)) {
return { lines, anchorCount, contextPairsCount };
}
contextPairsCount++;
}
}
const unit = buildEvidenceUnit(l0, l1ByFloor);
units.push(unit);
usedL0Ids.add(l0.id);
}
return { lines, anchorCount, contextPairsCount };
// 按楼层排序
units.sort((a, b) => a.l0.floor - b.l0.floor);
return units;
}
// ─────────────────────────────────────────────────────────────────────────────
// 事件格式化（L2→L0→L1 层级）
// ─────────────────────────────────────────────────────────────────────────────
/**
* 格式化事件（含 L0→L1 证据）
* @param {object} eventItem - 事件召回项
* @param {number} idx - 编号
* @param {EvidenceUnit[]} evidenceUnits - 该事件的证据单元
* @param {Map<string, object>} causalById - 因果事件索引
* @returns {string} 格式化后的文本
*/
function formatEventWithEvidence(eventItem, idx, evidenceUnits, causalById) {
const ev = eventItem.event || {};
const time = ev.timeLabel || "";
const title = String(ev.title || "").trim();
const people = (ev.participants || []).join(" / ").trim();
const summary = cleanSummary(ev.summary);
const displayTitle = title || people || ev.id || "事件";
const header = time ? `${idx}.【${time}】${displayTitle}` : `${idx}. ${displayTitle}`;
const lines = [header];
if (people && displayTitle !== people) lines.push(` ${people}`);
lines.push(` ${summary}`);
// 因果链
for (const cid of ev.causedBy || []) {
const c = causalById?.get(cid);
if (c) lines.push(formatCausalEventLine(c));
}
// L0→L1 证据单元
for (const unit of evidenceUnits) {
lines.push(...formatEvidenceUnit(unit));
}
return lines.join("\n");
}
// ─────────────────────────────────────────────────────────────────────────────
@@ -581,19 +556,22 @@ export function buildNonVectorPromptText() {
* 构建向量模式注入文本
* @param {object} store - 存储对象
* @param {object} recallResult - 召回结果
* @param {Map} causalById - 因果事件索引
* @param {Map<string, object>} causalById - 因果事件索引
* @param {string[]} focusEntities - 焦点实体
* @param {object} meta - 元数据
* @param {object} metrics - 指标对象
* @returns {Promise<{promptText: string, injectionLogText: string, injectionStats: object, metrics: object}>}
* @returns {Promise<{promptText: string, injectionStats: object, metrics: object}>}
*/
async function buildVectorPrompt(store, recallResult, causalById, focusEntities = [], meta = null, metrics = null) {
async function buildVectorPrompt(store, recallResult, causalById, focusEntities, meta, metrics) {
const T_Start = performance.now();
const { chatId } = getContext();
const data = store.json || {};
const total = { used: 0, max: MAIN_BUDGET_MAX };
// 从 recallResult 解构
const l0Selected = recallResult?.l0Selected || [];
const l1ByFloor = recallResult?.l1ByFloor || new Map();
// 装配结果
const assembled = {
constraints: { lines: [], tokens: 0 },
@@ -610,15 +588,9 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
constraint: { count: 0, tokens: 0, filtered: 0 },
arc: { count: 0, tokens: 0 },
event: { selected: 0, tokens: 0 },
evidence: { attached: 0, tokens: 0 },
distantEvidence: { injected: 0, tokens: 0, anchorCount: 0, contextPairs: 0 },
};
const recentEvidenceStats = {
injected: 0,
tokens: 0,
floorRange: "N/A",
contextPairs: 0,
evidence: { l0InEvents: 0, l1InEvents: 0, tokens: 0 },
distantEvidence: { units: 0, tokens: 0 },
recentEvidence: { units: 0, tokens: 0 },
};
const eventDetails = {
@@ -627,6 +599,9 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
relatedCount: 0,
};
// 已消费的 L0 ID 集合（事件区域消费后，evidence 区域不再重复）
const usedL0Ids = new Set();
// ═══════════════════════════════════════════════════════════════════════
// [Constraints] L3 Facts → 世界约束
// ═══════════════════════════════════════════════════════════════════════
@@ -698,70 +673,9 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
}
// ═══════════════════════════════════════════════════════════════════════
// [Events] L2 Events → 直接命中 + 相似命中 + 因果链
// [Events] L2 Events → 直接命中 + 相似命中 + 因果链 + L0→L1 证据
// ═══════════════════════════════════════════════════════════════════════
const eventHits = (recallResult?.events || []).filter(e => e?.event?.summary);
const evidenceChunks = recallResult?.evidenceChunks || [];
const usedChunkIds = new Set();
/**
* 为事件选择最佳证据 chunk
* @param {object} eventObj - 事件对象
* @returns {object|null} 最佳 chunk
*/
function pickBestEvidenceForEvent(eventObj) {
const range = parseFloorRange(eventObj?.summary);
if (!range) return null;
let best = null;
for (const c of evidenceChunks) {
if (usedChunkIds.has(c.chunkId)) continue;
if (c.floor < range.start || c.floor > range.end) continue;
if (!best) {
best = c;
} else if (c.isAnchorVirtual && !best.isAnchorVirtual) {
best = c;
} else if (c.isAnchorVirtual === best.isAnchorVirtual && (c.chunkIdx ?? 0) < (best.chunkIdx ?? 0)) {
best = c;
}
}
return best;
}
/**
* 格式化事件带证据
* @param {object} eventItem - 事件项
* @param {number} idx - 编号
* @param {object} chunk - 证据 chunk
* @returns {string} 格式化后的文本
*/
function formatEventWithEvidence(eventItem, idx, chunk) {
const ev = eventItem.event || {};
const time = ev.timeLabel || "";
const title = String(ev.title || "").trim();
const people = (ev.participants || []).join(" / ").trim();
const summary = cleanSummary(ev.summary);
const displayTitle = title || people || ev.id || "事件";
const header = time ? `${idx}.【${time}】${displayTitle}` : `${idx}. ${displayTitle}`;
const lines = [header];
if (people && displayTitle !== people) lines.push(` ${people}`);
lines.push(` ${summary}`);
for (const cid of ev.causedBy || []) {
const c = causalById?.get(cid);
if (c) lines.push(formatCausalEventLine(c, causalById));
}
if (chunk) {
lines.push(` ${formatEvidenceFullLine(chunk)}`);
}
return lines.join("\n");
}
const candidates = [...eventHits].sort((a, b) => (b.similarity || 0) - (a.similarity || 0));
@@ -775,52 +689,91 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
const isDirect = e._recallType === "DIRECT";
const bestChunk = pickBestEvidenceForEvent(e.event);
// 收集该事件范围内的 L0→L1 证据单元
const evidenceUnits = collectEvidenceForEvent(e.event, l0Selected, l1ByFloor, usedL0Ids);
let text = formatEventWithEvidence(e, 0, bestChunk);
let cost = estimateTokens(text);
let hasEvidence = !!bestChunk;
let chosenChunk = bestChunk || null;
// 格式化事件（含证据）
const text = formatEventWithEvidence(e, 0, evidenceUnits, causalById);
const cost = estimateTokens(text);
// 预算检查：整个事件（含证据）作为原子单元
if (total.used + cost > total.max) {
text = formatEventWithEvidence(e, 0, null);
cost = estimateTokens(text);
hasEvidence = false;
chosenChunk = null;
// 尝试不带证据的版本
const textNoEvidence = formatEventWithEvidence(e, 0, [], causalById);
const costNoEvidence = estimateTokens(textNoEvidence);
if (total.used + cost > total.max) {
if (total.used + costNoEvidence > total.max) {
continue;
}
// 放入不带证据的版本，归还已消费的 L0 ID
for (const unit of evidenceUnits) {
usedL0Ids.delete(unit.l0.id);
}
if (isDirect) {
selectedDirect.push({
event: e.event, text: textNoEvidence, tokens: costNoEvidence,
evidenceUnits: [], candidateRank,
});
} else {
selectedRelated.push({
event: e.event, text: textNoEvidence, tokens: costNoEvidence,
evidenceUnits: [], candidateRank,
});
}
injectionStats.event.selected++;
injectionStats.event.tokens += costNoEvidence;
total.used += costNoEvidence;
eventDetails.list.push({
title: e.event?.title || e.event?.id,
isDirect,
hasEvidence: false,
tokens: costNoEvidence,
similarity: e.similarity || 0,
l0Count: 0,
l1Count: 0,
});
continue;
}
// 预算充足，放入完整版本
const l0Count = evidenceUnits.length;
let l1Count = 0;
for (const unit of evidenceUnits) {
if (unit.userL1) l1Count++;
if (unit.aiL1) l1Count++;
}
if (isDirect) {
selectedDirect.push({ event: e.event, text, tokens: cost, chunk: chosenChunk, hasEvidence, candidateRank });
selectedDirect.push({
event: e.event, text, tokens: cost,
evidenceUnits, candidateRank,
});
} else {
selectedRelated.push({ event: e.event, text, tokens: cost, chunk: chosenChunk, hasEvidence, candidateRank });
selectedRelated.push({
event: e.event, text, tokens: cost,
evidenceUnits, candidateRank,
});
}
injectionStats.event.selected++;
injectionStats.event.tokens += cost;
injectionStats.evidence.l0InEvents += l0Count;
injectionStats.evidence.l1InEvents += l1Count;
total.used += cost;
if (hasEvidence && bestChunk) {
const chunkLine = formatEvidenceFullLine(bestChunk);
const ct = estimateTokens(chunkLine);
injectionStats.evidence.attached++;
injectionStats.evidence.tokens += ct;
usedChunkIds.add(bestChunk.chunkId);
injectionStats.event.tokens += Math.max(0, cost - ct);
} else {
injectionStats.event.tokens += cost;
}
eventDetails.list.push({
title: e.event?.title || e.event?.id,
isDirect,
hasEvidence,
hasEvidence: l0Count > 0,
tokens: cost,
similarity: e.similarity || 0,
hasAnchorEvidence: bestChunk?.isAnchorVirtual || false,
l0Count,
l1Count,
});
}
@@ -845,110 +798,81 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
assembled.relatedEvents.lines = relatedEventTexts;
// ═══════════════════════════════════════════════════════════════════════
// [Evidence - Distant] L1 Chunks → 远期证据（已总结范围）
// [Evidence - Distant] 远期证据（已总结范围，未被事件消费的 L0→L1）
// ═══════════════════════════════════════════════════════════════════════
const lastSummarized = store.lastSummarizedMesId ?? -1;
const lastChunkFloor = meta?.lastChunkFloor ?? -1;
const keepVisible = store.keepVisibleCount ?? 3;
const distantContextFloors = new Set();
const distantCandidates = evidenceChunks
.filter(c => !usedChunkIds.has(c.chunkId))
.filter(c => c.floor <= lastSummarized);
// 收集未被事件消费的 L0，按 rerankScore 降序
const remainingL0 = l0Selected
.filter(l0 => !usedL0Ids.has(l0.id))
.sort((a, b) => (b.rerankScore || 0) - (a.rerankScore || 0));
for (const c of distantCandidates) {
if (c.isAnchorVirtual) continue;
const pairFloor = getContextFloor(c);
if (pairFloor >= 0) distantContextFloors.add(pairFloor);
}
// 远期：floor <= lastSummarized
const distantL0 = remainingL0.filter(l0 => l0.floor <= lastSummarized);
let contextChunksByFloor = new Map();
if (chatId && distantContextFloors.size > 0) {
try {
const contextChunks = await getChunksByFloors(chatId, Array.from(distantContextFloors));
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 (distantCandidates.length && total.used < total.max) {
if (distantL0.length && total.used < total.max) {
const distantBudget = { used: 0, max: Math.min(DISTANT_EVIDENCE_MAX, total.max - total.used) };
const result = assembleEvidenceByFloor(
distantCandidates.sort((a, b) => (a.floor - b.floor) || ((a.chunkIdx ?? 0) - (b.chunkIdx ?? 0))),
contextChunksByFloor,
distantBudget
);
// 按楼层排序（时间顺序）
distantL0.sort((a, b) => a.floor - b.floor);
for (const l0 of distantL0) {
const unit = buildEvidenceUnit(l0, l1ByFloor);
// 原子单元预算检查
if (distantBudget.used + unit.totalTokens > distantBudget.max) continue;
const unitLines = formatEvidenceUnit(unit);
for (const line of unitLines) {
assembled.distantEvidence.lines.push(line);
}
distantBudget.used += unit.totalTokens;
usedL0Ids.add(l0.id);
injectionStats.distantEvidence.units++;
}
assembled.distantEvidence.lines = result.lines;
assembled.distantEvidence.tokens = distantBudget.used;
total.used += distantBudget.used;
injectionStats.distantEvidence.injected = result.lines.length;
injectionStats.distantEvidence.tokens = distantBudget.used;
injectionStats.distantEvidence.anchorCount = result.anchorCount;
injectionStats.distantEvidence.contextPairs = result.contextPairsCount;
}
// ═══════════════════════════════════════════════════════════════════════
// [Evidence - Recent] L1 Chunks → 近期证据（未总结范围，独立预算）
// [Evidence - Recent] 近期证据（未总结范围，独立预算）
// ═══════════════════════════════════════════════════════════════════════
const recentStart = lastSummarized + 1;
const recentEnd = lastChunkFloor - keepVisible;
if (evidenceChunks.length && recentEnd >= recentStart) {
const recentCandidates = evidenceChunks
.filter(c => !usedChunkIds.has(c.chunkId))
.filter(c => c.floor >= recentStart && c.floor <= recentEnd);
if (recentEnd >= recentStart) {
const recentL0 = remainingL0
.filter(l0 => !usedL0Ids.has(l0.id))
.filter(l0 => l0.floor >= recentStart && l0.floor <= recentEnd);
const recentContextFloors = new Set();
for (const c of recentCandidates) {
if (c.isAnchorVirtual) continue;
const pairFloor = getContextFloor(c);
if (pairFloor >= 0) recentContextFloors.add(pairFloor);
}
if (chatId && recentContextFloors.size > 0) {
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);
}
}
}
if (recentCandidates.length) {
if (recentL0.length) {
const recentBudget = { used: 0, max: RECENT_EVIDENCE_MAX };
const result = assembleEvidenceByFloor(
recentCandidates.sort((a, b) => (a.floor - b.floor) || ((a.chunkIdx ?? 0) - (b.chunkIdx ?? 0))),
contextChunksByFloor,
recentBudget
);
// 按楼层排序（时间顺序）
recentL0.sort((a, b) => a.floor - b.floor);
for (const l0 of recentL0) {
const unit = buildEvidenceUnit(l0, l1ByFloor);
if (recentBudget.used + unit.totalTokens > recentBudget.max) continue;
const unitLines = formatEvidenceUnit(unit);
for (const line of unitLines) {
assembled.recentEvidence.lines.push(line);
}
recentBudget.used += unit.totalTokens;
usedL0Ids.add(l0.id);
injectionStats.recentEvidence.units++;
}
assembled.recentEvidence.lines = result.lines;
assembled.recentEvidence.tokens = recentBudget.used;
recentEvidenceStats.injected = result.lines.length;
recentEvidenceStats.tokens = recentBudget.used;
recentEvidenceStats.floorRange = `${recentStart + 1}~${recentEnd + 1}楼`;
recentEvidenceStats.contextPairs = result.contextPairsCount;
injectionStats.recentEvidence.tokens = recentBudget.used;
}
}
@@ -984,7 +908,7 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
metrics.timing.evidenceAssembly = Math.round(performance.now() - T_Start - (metrics.timing.constraintFilter || 0));
metrics.timing.formatting = 0;
}
return { promptText: "", injectionLogText: "", injectionStats, metrics };
return { promptText: "", injectionStats, metrics };
}
const promptText =
@@ -1009,15 +933,16 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
metrics.budget.utilization = Math.round(metrics.budget.total / TOTAL_BUDGET_MAX * 100);
metrics.budget.breakdown = {
constraints: assembled.constraints.tokens,
events: injectionStats.event.tokens + injectionStats.evidence.tokens,
events: injectionStats.event.tokens,
distantEvidence: injectionStats.distantEvidence.tokens,
recentEvidence: recentEvidenceStats.tokens || 0,
recentEvidence: injectionStats.recentEvidence.tokens,
arcs: assembled.arcs.tokens,
};
metrics.evidence.tokens = injectionStats.distantEvidence.tokens + (recentEvidenceStats.tokens || 0);
metrics.evidence.contextPairsAdded = injectionStats.distantEvidence.contextPairs + recentEvidenceStats.contextPairs;
metrics.evidence.assemblyTime = Math.round(performance.now() - T_Start - (metrics.timing.constraintFilter || 0) - metrics.formatting.time);
metrics.evidence.tokens = injectionStats.distantEvidence.tokens + injectionStats.recentEvidence.tokens;
metrics.evidence.assemblyTime = Math.round(
performance.now() - T_Start - (metrics.timing.constraintFilter || 0) - metrics.formatting.time
);
metrics.timing.evidenceAssembly = metrics.evidence.assemblyTime;
const totalFacts = allFacts.length;
@@ -1026,76 +951,19 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
: 100;
metrics.quality.eventPrecisionProxy = metrics.event?.similarityDistribution?.mean || 0;
const totalSelected = metrics.evidence.selected || 0;
const attached = injectionStats.evidence.attached;
metrics.quality.evidenceDensity = totalSelected > 0
? Math.round(attached / totalSelected * 100)
const totalL0Selected = l0Selected.length;
const l0WithL1 = l0Selected.filter(l0 => {
const pair = l1ByFloor.get(l0.floor);
return pair?.aiTop1 || pair?.userTop1;
}).length;
metrics.quality.l1AttachRate = totalL0Selected > 0
? Math.round(l0WithL1 / totalL0Selected * 100)
: 0;
const selectedReal = metrics.evidence.selectedByType?.chunkReal || 0;
const selectedTotal = metrics.evidence.selected || 0;
metrics.quality.chunkRealRatio = selectedTotal > 0
? Math.round(selectedReal / selectedTotal * 100)
: 0;
metrics.quality.potentialIssues = detectIssues(metrics);
}
return { promptText, injectionLogText: "", injectionStats, metrics };
}
// ─────────────────────────────────────────────────────────────────────────────
// 因果证据补充
// ─────────────────────────────────────────────────────────────────────────────
/**
* 为因果事件附加证据
* @param {object[]} causalChain - 因果链
* @param {Map} eventVectorMap - 事件向量索引
* @param {Map} chunkVectorMap - chunk 向量索引
* @param {Map} chunksMap - chunks 索引
*/
async function attachEvidenceToCausalEvents(causalChain, eventVectorMap, chunkVectorMap, chunksMap) {
for (const c of causalChain) {
c._evidenceChunk = null;
const ev = c.event;
if (!ev?.id) continue;
const evVec = eventVectorMap.get(ev.id);
if (!evVec?.length) continue;
const range = parseFloorRange(ev.summary);
if (!range) continue;
const candidateChunks = [];
for (const [chunkId, chunk] of chunksMap) {
if (chunk.floor >= range.start && chunk.floor <= range.end) {
const vec = chunkVectorMap.get(chunkId);
if (vec?.length) candidateChunks.push({ chunk, vec });
}
}
if (!candidateChunks.length) continue;
let best = null;
let bestSim = -1;
for (const { chunk, vec } of candidateChunks) {
const sim = cosineSimilarity(evVec, vec);
if (sim > bestSim) {
bestSim = sim;
best = chunk;
}
}
if (best && bestSim > 0.3) {
c._evidenceChunk = {
floor: best.floor,
speaker: best.speaker,
text: best.text,
similarity: bestSim,
};
}
}
return { promptText, injectionStats, metrics };
}
// ─────────────────────────────────────────────────────────────────────────────
@@ -1150,43 +1018,16 @@ export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
recallResult = {
...recallResult,
events: recallResult?.events || [],
evidenceChunks: recallResult?.evidenceChunks || [],
l0Selected: recallResult?.l0Selected || [],
l1ByFloor: recallResult?.l1ByFloor || new Map(),
causalChain: recallResult?.causalChain || [],
focusEntities: recallResult?.focusEntities || [],
logText: recallResult?.logText || "",
metrics: recallResult?.metrics || null,
};
const causalChain = recallResult.causalChain || [];
if (causalChain.length > 0) {
if (chatId) {
try {
const floors = new Set();
for (const c of causalChain) {
const r = parseFloorRange(c.event?.summary);
if (!r) continue;
for (let f = r.start; f <= r.end; f++) floors.add(f);
}
const [chunksList, chunkVecs, eventVecs] = await Promise.all([
getChunksByFloors(chatId, Array.from(floors)),
getAllChunkVectors(chatId),
getAllEventVectors(chatId),
]);
const chunksMap = new Map(chunksList.map(c => [c.chunkId, c]));
const chunkVectorMap = new Map(chunkVecs.map(v => [v.chunkId, v.vector]));
const eventVectorMap = new Map(eventVecs.map(v => [v.eventId, v.vector]));
await attachEvidenceToCausalEvents(causalChain, eventVectorMap, chunkVectorMap, chunksMap);
} catch (e) {
xbLog.warn(MODULE_ID, "Causal evidence attachment failed", e);
}
}
}
// 构建因果事件索引
causalById = new Map(
recallResult.causalChain
(recallResult.causalChain || [])
.map(c => [c?.event?.id, c])
.filter(x => x[0])
);
@@ -1210,7 +1051,7 @@ export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
const hasUseful =
(recallResult?.events?.length || 0) > 0 ||
(recallResult?.evidenceChunks?.length || 0) > 0 ||
(recallResult?.l0Selected?.length || 0) > 0 ||
(recallResult?.causalChain?.length || 0) > 0;
if (!hasUseful) {