975 lines
42 KiB
JavaScript
975 lines
42 KiB
JavaScript
// Story Summary - Prompt Injection
|
||
// 注入格式：世界状态 → 亲身经历(含闪回) → 相关背景(含闪回) → 记忆碎片 → 人物弧光
|
||
|
||
import { getContext } from "../../../../../../extensions.js";
|
||
import {
|
||
extension_prompts,
|
||
extension_prompt_types,
|
||
extension_prompt_roles,
|
||
} from "../../../../../../../script.js";
|
||
import { xbLog } from "../../../core/debug-core.js";
|
||
import { getSummaryStore } from "../data/store.js";
|
||
import { getVectorConfig, getSummaryPanelConfig, getSettings } from "../data/config.js";
|
||
import { recallMemory, buildQueryText } from "../vector/recall.js";
|
||
import { getChunksByFloors, getAllChunkVectors, getAllEventVectors } from "../vector/chunk-store.js";
|
||
|
||
const MODULE_ID = "summaryPrompt";
|
||
const SUMMARY_PROMPT_KEY = "LittleWhiteBox_StorySummary";
|
||
|
||
// 预算：只保留主预算与 L3 上限，其它由装配算法决定
|
||
const BUDGET = { total: 10000, l3Max: 2000 };
|
||
|
||
// 你确认的参数
|
||
const TARGET_UTILIZATION = 0.8;
|
||
const TOP_RELEVANCE_COUNT = 5;
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// Injection log
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
function pct(n, d) {
|
||
return d > 0 ? Math.round((n / d) * 100) : 0;
|
||
}
|
||
|
||
function formatInjectionLog(stats) {
|
||
const lines = [
|
||
"",
|
||
"╔══════════════════════════════════════════════════════════════╗",
|
||
"║ Prompt Injection Report ║",
|
||
"╠══════════════════════════════════════════════════════════════╣",
|
||
`║ Token budget: ${stats.budget.max}`,
|
||
"╚══════════════════════════════════════════════════════════════╝",
|
||
"",
|
||
];
|
||
|
||
lines.push("┌─────────────────────────────────────────────────────────────┐");
|
||
lines.push("│ [Packing] Budget-aware assembly │");
|
||
lines.push("└─────────────────────────────────────────────────────────────┘");
|
||
if (stats.packing) {
|
||
lines.push(
|
||
` Target utilization: ${(stats.packing.targetUtilization * 100).toFixed(0)}%`
|
||
);
|
||
lines.push(
|
||
` L2 budget: ${stats.packing.l2Used} / ${stats.packing.l2Max} (${pct(stats.packing.l2Used, stats.packing.l2Max)}%)`
|
||
);
|
||
lines.push(
|
||
` Selected events: ${stats.packing.selectedEvents} (DIRECT: ${stats.packing.selectedDirect}, SIMILAR: ${stats.packing.selectedSimilar})`
|
||
);
|
||
lines.push(
|
||
` Evidence levels: E3=${stats.packing.e3} | E2=${stats.packing.e2} | E1=${stats.packing.e1} | E0=${stats.packing.e0}`
|
||
);
|
||
lines.push(` Evidence chunks (total): ${stats.packing.evidenceChunks}`);
|
||
} else {
|
||
lines.push(" (no packing stats)");
|
||
}
|
||
lines.push("");
|
||
|
||
lines.push("┌─────────────────────────────────────────────────────────────┐");
|
||
lines.push("│ [World] L3 │");
|
||
lines.push("└─────────────────────────────────────────────────────────────┘");
|
||
lines.push(` Injected: ${stats.world.count} | Tokens: ${stats.world.tokens}`);
|
||
lines.push("");
|
||
|
||
lines.push("┌─────────────────────────────────────────────────────────────┐");
|
||
lines.push("│ [Direct] │");
|
||
lines.push("└─────────────────────────────────────────────────────────────┘");
|
||
lines.push(
|
||
` Events: ${stats.direct.recalled} -> ${stats.direct.injected}${stats.direct.recalled > stats.direct.injected ? ` (budget cut ${stats.direct.recalled - stats.direct.injected})` : ""}`
|
||
);
|
||
lines.push(` Causal: ${stats.direct.causalCount}`);
|
||
lines.push(` L1 chunks: ${stats.direct.chunksCount}`);
|
||
lines.push(` Tokens: ${stats.direct.tokens}`);
|
||
lines.push("");
|
||
|
||
lines.push("┌─────────────────────────────────────────────────────────────┐");
|
||
lines.push("│ [Similar] │");
|
||
lines.push("└─────────────────────────────────────────────────────────────┘");
|
||
lines.push(
|
||
` Events: ${stats.similar.recalled} -> ${stats.similar.injected}${stats.similar.recalled > stats.similar.injected ? ` (budget cut ${stats.similar.recalled - stats.similar.injected})` : ""}`
|
||
);
|
||
lines.push(` Causal: ${stats.similar.causalCount}`);
|
||
lines.push(` L1 chunks: ${stats.similar.chunksCount}`);
|
||
lines.push(` Tokens: ${stats.similar.tokens}`);
|
||
lines.push("");
|
||
|
||
lines.push("┌─────────────────────────────────────────────────────────────┐");
|
||
lines.push("│ [Orphans] │");
|
||
lines.push("└─────────────────────────────────────────────────────────────┘");
|
||
lines.push(
|
||
` Chunks: ${stats.orphans.recalled} -> ${stats.orphans.injected}${stats.orphans.recalled > stats.orphans.injected ? ` (budget cut ${stats.orphans.recalled - stats.orphans.injected})` : ""}`
|
||
);
|
||
lines.push(` Tokens: ${stats.orphans.tokens}`);
|
||
lines.push("");
|
||
|
||
lines.push("┌─────────────────────────────────────────────────────────────┐");
|
||
lines.push("│ [Arcs] │");
|
||
lines.push("└─────────────────────────────────────────────────────────────┘");
|
||
lines.push(` Injected: ${stats.arcs.count} | Tokens: ${stats.arcs.tokens}`);
|
||
lines.push("");
|
||
|
||
lines.push("┌─────────────────────────────────────────────────────────────┐");
|
||
lines.push("│ [Total] │");
|
||
lines.push("└─────────────────────────────────────────────────────────────┘");
|
||
lines.push(
|
||
` Tokens: ${stats.budget.used} / ${stats.budget.max} (${Math.round((stats.budget.used / stats.budget.max) * 100)}%)`
|
||
);
|
||
lines.push("");
|
||
|
||
return lines.join("\n");
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// 向量工具
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
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;
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// 工具函数
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
function estimateTokens(text) {
|
||
if (!text) return 0;
|
||
const s = String(text);
|
||
const zh = (s.match(/[\u4e00-\u9fff]/g) || []).length;
|
||
return Math.ceil(zh + (s.length - zh) / 4);
|
||
}
|
||
|
||
function clamp(n, min, max) {
|
||
return Math.max(min, Math.min(max, n));
|
||
}
|
||
|
||
function pushWithBudget(lines, text, state) {
|
||
const t = estimateTokens(text);
|
||
if (state.used + t > state.max) return false;
|
||
lines.push(text);
|
||
state.used += t;
|
||
return true;
|
||
}
|
||
|
||
// 从 summary 解析楼层范围：(#321-322) 或 (#321)
|
||
function parseFloorRange(summary) {
|
||
if (!summary) return null;
|
||
const match = String(summary).match(/\(#(\d+)(?:-(\d+))?\)/);
|
||
if (!match) return null;
|
||
|
||
// summary 里写的是 #楼层（1-based），chunks 里 floor 是消息下标（0-based）
|
||
const start = Math.max(0, parseInt(match[1], 10) - 1);
|
||
const end = Math.max(
|
||
0,
|
||
(match[2] ? parseInt(match[2], 10) : parseInt(match[1], 10)) - 1
|
||
);
|
||
|
||
return { start, end };
|
||
}
|
||
|
||
// 去掉 summary 末尾的楼层标记
|
||
function cleanSummary(summary) {
|
||
return String(summary || "")
|
||
.replace(/\s*\(#\d+(?:-\d+)?\)\s*$/, "")
|
||
.trim();
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// Evidence Windowing (证据窗口)
|
||
// E1: 核心1条 / E2: ±1(约3条) / E3: ±2(约5条)
|
||
// 不改chunk切分，不做重叠，只在注入时补邻域，提高语义完整度。
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
const EVIDENCE_LEVEL = {
|
||
E0: 0,
|
||
E1: 1,
|
||
E2: 2,
|
||
E3: 3,
|
||
};
|
||
|
||
function getEvidenceWindowRadius(level) {
|
||
if (level === EVIDENCE_LEVEL.E3) return 2;
|
||
if (level === EVIDENCE_LEVEL.E2) return 1;
|
||
return 0;
|
||
}
|
||
|
||
function buildChunksByFloorMap(chunks) {
|
||
const map = new Map();
|
||
for (const c of chunks || []) {
|
||
const f = c.floor;
|
||
if (!map.has(f)) map.set(f, []);
|
||
map.get(f).push(c);
|
||
}
|
||
for (const arr of map.values()) {
|
||
arr.sort((a, b) => (a.chunkIdx ?? 0) - (b.chunkIdx ?? 0));
|
||
}
|
||
return map;
|
||
}
|
||
|
||
function pickAnchorChunkIdx(eventItem, floorChunks, recalledChunksInRange = []) {
|
||
// 优先：用本轮召回的chunks里，同楼层且相似度最高的作为anchor
|
||
let best = null;
|
||
for (const rc of recalledChunksInRange) {
|
||
if (rc.floor !== eventItem._evidenceFloor) continue;
|
||
if (!best || (rc.similarity || 0) > (best.similarity || 0)) best = rc;
|
||
}
|
||
if (best && best.chunkIdx != null) return best.chunkIdx;
|
||
|
||
// 退化：该楼层第一个chunk
|
||
const first = floorChunks?.[0];
|
||
return first?.chunkIdx ?? 0;
|
||
}
|
||
|
||
function getEvidenceChunksForEvent(eventItem, chunksByFloor, recalledChunksInRange, evidenceLevel) {
|
||
if (evidenceLevel === EVIDENCE_LEVEL.E0) return [];
|
||
|
||
const floor = eventItem._evidenceFloor;
|
||
const floorChunks = chunksByFloor.get(floor) || [];
|
||
if (!floorChunks.length) return [];
|
||
|
||
const radius = getEvidenceWindowRadius(evidenceLevel);
|
||
const anchorIdx = pickAnchorChunkIdx(eventItem, floorChunks, recalledChunksInRange);
|
||
|
||
// 找到anchor在floorChunks中的位置
|
||
const pos = floorChunks.findIndex(c => (c.chunkIdx ?? 0) === anchorIdx);
|
||
const anchorPos = pos >= 0 ? pos : 0;
|
||
|
||
const start = clamp(anchorPos - radius, 0, floorChunks.length - 1);
|
||
const end = clamp(anchorPos + radius, 0, floorChunks.length - 1);
|
||
const selected = floorChunks.slice(start, end + 1);
|
||
|
||
// E1只取核心一条
|
||
if (evidenceLevel === EVIDENCE_LEVEL.E1) return [floorChunks[anchorPos]];
|
||
return selected;
|
||
}
|
||
|
||
function downgrade(level) {
|
||
if (level <= EVIDENCE_LEVEL.E0) return EVIDENCE_LEVEL.E0;
|
||
return level - 1;
|
||
}
|
||
|
||
function chooseInitialEvidenceLevel(e, isTop) {
|
||
if (isTop) return EVIDENCE_LEVEL.E3;
|
||
if (e._recallType === "DIRECT") return EVIDENCE_LEVEL.E2;
|
||
return EVIDENCE_LEVEL.E1;
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// L1 → L2 归属：这里只挂“候选chunks”，最终证据窗口在装配阶段决定
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
function attachChunksToEvents(events, chunks) {
|
||
const usedChunkIds = new Set();
|
||
|
||
for (const e of events) {
|
||
e._candidateChunks = [];
|
||
const range = parseFloorRange(e.event?.summary);
|
||
if (!range) continue;
|
||
|
||
for (const c of chunks) {
|
||
if (c.floor >= range.start && c.floor <= range.end) {
|
||
if (!usedChunkIds.has(c.chunkId)) {
|
||
e._candidateChunks.push(c);
|
||
usedChunkIds.add(c.chunkId);
|
||
}
|
||
}
|
||
}
|
||
|
||
e._candidateChunks.sort(
|
||
(a, b) => (a.floor - b.floor) || ((b.similarity || 0) - (a.similarity || 0))
|
||
);
|
||
}
|
||
|
||
const orphans = chunks
|
||
.filter(c => !usedChunkIds.has(c.chunkId))
|
||
.sort((a, b) => (b.similarity || 0) - (a.similarity || 0));
|
||
|
||
return { events, orphans };
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// 因果事件证据补充：用 eventVector 匹配最相关的 chunk
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
async function attachEvidenceToaCausalEvents(causalEvents, eventVectorMap, chunkVectorMap, chunksMap) {
|
||
for (const c of causalEvents) {
|
||
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,
|
||
};
|
||
}
|
||
}
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// 格式化函数
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
function formatWorldLines(world) {
|
||
return [...(world || [])]
|
||
.sort((a, b) => (b.floor || 0) - (a.floor || 0))
|
||
.map(w => `- ${w.topic}：${w.content}`);
|
||
}
|
||
|
||
function formatChunkLine(c) {
|
||
const text = String(c.text || "");
|
||
const preview = text.length > 80 ? text.slice(0, 80) + "..." : text;
|
||
const speaker = c.isUser ? "{{user}}" : "{{char}}";
|
||
return `› #${c.floor + 1} [${speaker}] ${preview}`;
|
||
}
|
||
|
||
function formatEventBlock(e, idx, isHighRelevance = false, evidenceChunks = []) {
|
||
const ev = e.event || {};
|
||
const time = ev.timeLabel || "";
|
||
const title = String(ev.title || "").trim();
|
||
const people = (ev.participants || []).join(" / ").trim();
|
||
const summary = cleanSummary(ev.summary);
|
||
|
||
const lines = [];
|
||
|
||
const displayTitle = title || people || ev.id || "事件";
|
||
const marker = isHighRelevance ? "★" : "";
|
||
const header = time ? `${marker}${idx}.【${time}】${displayTitle}` : `${marker}${idx}. ${displayTitle}`;
|
||
lines.push(header);
|
||
|
||
if (people && displayTitle !== people) {
|
||
lines.push(` ${people}`);
|
||
}
|
||
|
||
lines.push(` ${summary}`);
|
||
|
||
for (const c of evidenceChunks || []) {
|
||
lines.push(` ${formatChunkLine(c)}`);
|
||
}
|
||
|
||
return lines.join("\n");
|
||
}
|
||
|
||
function formatCausalEventLine(causalItem, _causalById) {
|
||
const ev = causalItem?.event || {};
|
||
const depth = Math.max(1, Math.min(9, causalItem?._causalDepth || 1));
|
||
const indent = " │" + " ".repeat(depth - 1);
|
||
const prefix = `${indent}├─ 前因`;
|
||
|
||
const time = ev.timeLabel ? `【${ev.timeLabel}】` : "";
|
||
const people = (ev.participants || []).join(" / ");
|
||
const summary = cleanSummary(ev.summary);
|
||
|
||
const r = parseFloorRange(ev.summary);
|
||
const floorHint = r ? `(#${r.start + 1}${r.end !== r.start ? `-${r.end + 1}` : ""})` : "";
|
||
|
||
const lines = [];
|
||
lines.push(`${prefix}${time}${people ? ` ${people}` : ""}`);
|
||
const body = `${summary}${floorHint ? ` ${floorHint}` : ""}`.trim();
|
||
lines.push(`${indent} ${body}`);
|
||
|
||
const evidence = causalItem._evidenceChunk;
|
||
if (evidence) {
|
||
const speaker = evidence.speaker || "角色";
|
||
const preview = evidence.text.length > 60 ? evidence.text.slice(0, 60) + "..." : evidence.text;
|
||
lines.push(`${indent} › #${evidence.floor + 1} [${speaker}] ${preview}`);
|
||
}
|
||
|
||
return lines.join("\n");
|
||
}
|
||
|
||
function formatArcLine(a) {
|
||
const moments = (a.moments || [])
|
||
.map(m => (typeof m === "string" ? m : m.text))
|
||
.filter(Boolean);
|
||
|
||
if (moments.length) {
|
||
return `- ${a.name}：${moments.join(" → ")}（当前：${a.trajectory}）`;
|
||
}
|
||
return `- ${a.name}：${a.trajectory}`;
|
||
}
|
||
|
||
function buildTopKIdSet(directEvents, similarEvents) {
|
||
return new Set(
|
||
[...directEvents, ...similarEvents]
|
||
.sort((a, b) => (b.similarity || 0) - (a.similarity || 0))
|
||
.slice(0, TOP_RELEVANCE_COUNT)
|
||
.map(e => e.event?.id)
|
||
.filter(Boolean)
|
||
);
|
||
}
|
||
|
||
function computeEventTextCost(e, isTop, evidenceChunks = []) {
|
||
const tmp = formatEventBlock(e, 1, isTop, evidenceChunks);
|
||
return estimateTokens(tmp);
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// 非向量模式：沿用旧行为（简单、快）
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
function buildMemoryPromptVectorDisabled(store) {
|
||
const data = store.json || {};
|
||
const sections = [];
|
||
|
||
if (data.world?.length) {
|
||
const lines = formatWorldLines(data.world);
|
||
sections.push(`[世界状态] 请严格遵守\n${lines.join("\n")}`);
|
||
}
|
||
|
||
if (data.events?.length) {
|
||
const lines = data.events.map((ev, i) => {
|
||
const time = ev.timeLabel || "";
|
||
const title = ev.title || "";
|
||
const people = (ev.participants || []).join(" / ");
|
||
const summary = cleanSummary(ev.summary);
|
||
const header = time ? `${i + 1}.【${time}】${title || people}` : `${i + 1}. ${title || people}`;
|
||
return `${header}\n ${summary}`;
|
||
});
|
||
sections.push(`[剧情记忆]\n\n${lines.join("\n\n")}`);
|
||
}
|
||
|
||
if (data.arcs?.length) {
|
||
const lines = data.arcs.map(formatArcLine);
|
||
sections.push(`[人物弧光]\n${lines.join("\n")}`);
|
||
}
|
||
|
||
if (!sections.length) return { promptText: "", injectionLogText: "", injectionStats: null };
|
||
return {
|
||
promptText: `<剧情记忆>\n\n${sections.join("\n\n")}\n\n</剧情记忆>`,
|
||
injectionLogText: "",
|
||
injectionStats: null,
|
||
};
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// 预算驱动装配（向量模式）
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
async function buildMemoryPromptVectorEnabled(store, recallResult, causalById, queryEntities = []) {
|
||
const data = store.json || {};
|
||
const total = { used: 0, max: BUDGET.total };
|
||
const sections = [];
|
||
|
||
const injectionStats = {
|
||
budget: { max: BUDGET.total, used: 0 },
|
||
world: { count: 0, tokens: 0 },
|
||
direct: { recalled: 0, injected: 0, causalCount: 0, chunksCount: 0, tokens: 0 },
|
||
similar: { recalled: 0, injected: 0, causalCount: 0, chunksCount: 0, tokens: 0 },
|
||
orphans: { recalled: 0, injected: 0, tokens: 0 },
|
||
arcs: { count: 0, tokens: 0 },
|
||
packing: null,
|
||
};
|
||
|
||
const targetUsed = Math.floor(BUDGET.total * TARGET_UTILIZATION);
|
||
|
||
// [世界状态]
|
||
const worldLines = formatWorldLines(data.world);
|
||
if (worldLines.length) {
|
||
const l3 = { used: 0, max: Math.min(BUDGET.l3Max, total.max) };
|
||
const l3Lines = [];
|
||
|
||
for (const line of worldLines) {
|
||
if (!pushWithBudget(l3Lines, line, l3)) break;
|
||
}
|
||
|
||
if (l3Lines.length) {
|
||
sections.push(`[世界状态] 请严格遵守\n${l3Lines.join("\n")}`);
|
||
total.used += l3.used;
|
||
injectionStats.world.count = l3Lines.length;
|
||
injectionStats.world.tokens = l3.used;
|
||
}
|
||
}
|
||
|
||
// L1 → L2 归属
|
||
const events = recallResult?.events || [];
|
||
const chunks = recallResult?.chunks || [];
|
||
const { events: eventsWithChunks, orphans } = attachChunksToEvents(events, chunks);
|
||
|
||
const directEvents = eventsWithChunks.filter(e => e._recallType === "DIRECT");
|
||
const similarEvents = eventsWithChunks.filter(e => e._recallType !== "DIRECT");
|
||
|
||
injectionStats.direct.recalled = directEvents.length;
|
||
injectionStats.similar.recalled = similarEvents.length;
|
||
|
||
const topKIds = buildTopKIdSet(directEvents, similarEvents);
|
||
|
||
// 证据楼层选择：用事件 range.end 作为 evidenceFloor（贴近事件结尾）
|
||
const evidenceFloors = new Set();
|
||
for (const e of eventsWithChunks) {
|
||
const r = parseFloorRange(e.event?.summary);
|
||
if (!r) continue;
|
||
e._evidenceFloor = r.end;
|
||
evidenceFloors.add(r.end);
|
||
}
|
||
|
||
// 批量加载这些楼层 chunks，用于证据窗口
|
||
const { chatId } = getContext();
|
||
let chunksByFloor = new Map();
|
||
if (chatId && evidenceFloors.size) {
|
||
try {
|
||
const floorChunks = await getChunksByFloors(chatId, Array.from(evidenceFloors));
|
||
chunksByFloor = buildChunksByFloorMap(floorChunks);
|
||
} catch (e) {
|
||
xbLog.warn(MODULE_ID, "Failed to load floor chunks for evidence windowing", e);
|
||
}
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
// 预算装配（不再固定条数）
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
// L2预算：目标 65% 总预算，上限 80%（保守避免 L2 吞满全部）
|
||
const l2Target = Math.floor(BUDGET.total * 0.65);
|
||
const l2Ceil = Math.floor(BUDGET.total * 0.8);
|
||
const l2Budget = {
|
||
used: 0,
|
||
max: Math.min(l2Ceil, Math.max(0, BUDGET.total - total.used)),
|
||
};
|
||
|
||
const packStats = {
|
||
targetUtilization: TARGET_UTILIZATION,
|
||
l2Used: 0,
|
||
l2Max: l2Budget.max,
|
||
selectedEvents: 0,
|
||
selectedDirect: 0,
|
||
selectedSimilar: 0,
|
||
e3: 0,
|
||
e2: 0,
|
||
e1: 0,
|
||
e0: 0,
|
||
evidenceChunks: 0,
|
||
};
|
||
|
||
// 候选：按 similarity 降序（更贴近“本轮需要”）
|
||
const candidates = [...directEvents, ...similarEvents]
|
||
.filter(e => e?.event?.summary)
|
||
.sort((a, b) => (b.similarity || 0) - (a.similarity || 0));
|
||
|
||
const selected = []; // { e, evidenceLevel, evidenceChunks, cost, isTop }
|
||
const selectedIds = new Set();
|
||
|
||
for (const e of candidates) {
|
||
const id = e.event?.id;
|
||
if (!id || selectedIds.has(id)) continue;
|
||
|
||
const isTop = topKIds.has(id);
|
||
let level = chooseInitialEvidenceLevel(e, isTop);
|
||
|
||
const recalledInRange = e._candidateChunks || [];
|
||
|
||
// 从高到低降级，直到能塞入 L2 budget
|
||
while (true) {
|
||
const evChunks = getEvidenceChunksForEvent(e, chunksByFloor, recalledInRange, level);
|
||
const cost = computeEventTextCost(e, isTop, evChunks);
|
||
|
||
if (l2Budget.used + cost <= l2Budget.max) {
|
||
selected.push({ e, evidenceLevel: level, evidenceChunks: evChunks, cost, isTop });
|
||
selectedIds.add(id);
|
||
l2Budget.used += cost;
|
||
break;
|
||
}
|
||
|
||
if (level === EVIDENCE_LEVEL.E0) break;
|
||
level = downgrade(level);
|
||
}
|
||
|
||
// 达到 L2 目标就先停（后续仍可能做“证据升级”填预算）
|
||
if (l2Budget.used >= l2Target) break;
|
||
}
|
||
|
||
// 若总预算仍明显不足目标利用率，做一次“证据升级”填预算（安全填充）
|
||
if (total.used + l2Budget.used < targetUsed && l2Budget.used < l2Budget.max && selected.length) {
|
||
const upgradable = [...selected].sort((a, b) => {
|
||
if (a.isTop !== b.isTop) return a.isTop ? -1 : 1;
|
||
return (b.e.similarity || 0) - (a.e.similarity || 0);
|
||
});
|
||
|
||
for (const item of upgradable) {
|
||
if (total.used + l2Budget.used >= targetUsed) break;
|
||
if (l2Budget.used >= l2Budget.max) break;
|
||
|
||
const cur = item.evidenceLevel;
|
||
const next = cur >= EVIDENCE_LEVEL.E3 ? cur : cur + 1;
|
||
if (next === cur) continue;
|
||
|
||
const recalledInRange = item.e._candidateChunks || [];
|
||
const nextChunks = getEvidenceChunksForEvent(item.e, chunksByFloor, recalledInRange, next);
|
||
const nextCost = computeEventTextCost(item.e, item.isTop, nextChunks);
|
||
const delta = nextCost - item.cost;
|
||
|
||
if (delta <= 0) continue;
|
||
if (l2Budget.used + delta <= l2Budget.max) {
|
||
item.evidenceLevel = next;
|
||
item.evidenceChunks = nextChunks;
|
||
item.cost = nextCost;
|
||
l2Budget.used += delta;
|
||
}
|
||
}
|
||
}
|
||
|
||
// packing stats 汇总
|
||
packStats.l2Used = l2Budget.used;
|
||
|
||
for (const item of selected) {
|
||
packStats.selectedEvents++;
|
||
if (item.e._recallType === "DIRECT") packStats.selectedDirect++;
|
||
else packStats.selectedSimilar++;
|
||
|
||
if (item.evidenceLevel === EVIDENCE_LEVEL.E3) packStats.e3++;
|
||
else if (item.evidenceLevel === EVIDENCE_LEVEL.E2) packStats.e2++;
|
||
else if (item.evidenceLevel === EVIDENCE_LEVEL.E1) packStats.e1++;
|
||
else packStats.e0++;
|
||
|
||
packStats.evidenceChunks += item.evidenceChunks?.length || 0;
|
||
}
|
||
injectionStats.packing = packStats;
|
||
|
||
// 最终输出仍按时间线：按事件 summary 里的楼层范围 start 排序
|
||
function getEventFloorStart(ev) {
|
||
const r = parseFloorRange(ev?.summary);
|
||
return r?.start ?? Number.POSITIVE_INFINITY;
|
||
}
|
||
|
||
const selectedEventsOrdered = selected.sort(
|
||
(a, b) => getEventFloorStart(a.e.event) - getEventFloorStart(b.e.event)
|
||
);
|
||
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
// [亲身经历] DIRECT
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
{
|
||
const directLines = [];
|
||
let idx = 1;
|
||
let injectedCount = 0;
|
||
let causalCount = 0;
|
||
let chunksCount = 0;
|
||
|
||
for (const item of selectedEventsOrdered) {
|
||
if (item.e._recallType !== "DIRECT") continue;
|
||
|
||
const block = formatEventBlock(item.e, idx, item.isTop, item.evidenceChunks);
|
||
directLines.push(block);
|
||
injectedCount++;
|
||
chunksCount += item.evidenceChunks?.length || 0;
|
||
|
||
for (const cid of item.e.event?.causedBy || []) {
|
||
const c = causalById.get(cid);
|
||
if (!c) continue;
|
||
directLines.push(formatCausalEventLine(c, causalById));
|
||
causalCount++;
|
||
}
|
||
idx++;
|
||
}
|
||
|
||
if (directLines.length) {
|
||
const text = `[亲身经历]\n\n${directLines.join("\n\n")}`;
|
||
const t = estimateTokens(text);
|
||
if (total.used + t <= total.max) {
|
||
sections.push(text);
|
||
total.used += t;
|
||
injectionStats.direct.injected = injectedCount;
|
||
injectionStats.direct.causalCount = causalCount;
|
||
injectionStats.direct.chunksCount = chunksCount;
|
||
injectionStats.direct.tokens = t;
|
||
}
|
||
}
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
// [相关背景] SIMILAR
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
{
|
||
const similarLines = [];
|
||
let idx = (injectionStats.direct.injected || 0) + 1;
|
||
let injectedCount = 0;
|
||
let causalCount = 0;
|
||
let chunksCount = 0;
|
||
|
||
for (const item of selectedEventsOrdered) {
|
||
if (item.e._recallType === "DIRECT") continue;
|
||
|
||
const block = formatEventBlock(item.e, idx, item.isTop, item.evidenceChunks);
|
||
similarLines.push(block);
|
||
injectedCount++;
|
||
chunksCount += item.evidenceChunks?.length || 0;
|
||
|
||
for (const cid of item.e.event?.causedBy || []) {
|
||
const c = causalById.get(cid);
|
||
if (!c) continue;
|
||
similarLines.push(formatCausalEventLine(c, causalById));
|
||
causalCount++;
|
||
}
|
||
idx++;
|
||
}
|
||
|
||
if (similarLines.length) {
|
||
const text = `[相关背景]\n\n${similarLines.join("\n\n")}`;
|
||
const t = estimateTokens(text);
|
||
if (total.used + t <= total.max) {
|
||
sections.push(text);
|
||
total.used += t;
|
||
injectionStats.similar.injected = injectedCount;
|
||
injectionStats.similar.causalCount = causalCount;
|
||
injectionStats.similar.chunksCount = chunksCount;
|
||
injectionStats.similar.tokens = t;
|
||
}
|
||
}
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
// [记忆碎片] Orphans：按剩余预算自然装入（仍受预算约束），按时间排序
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
if (orphans.length && total.used < total.max) {
|
||
const l1 = { used: 0, max: total.max - total.used };
|
||
const lines = [];
|
||
|
||
injectionStats.orphans.recalled = orphans.length;
|
||
|
||
orphans.sort((a, b) => a.floor - b.floor);
|
||
|
||
for (const c of orphans) {
|
||
const line = formatChunkLine(c);
|
||
if (!pushWithBudget(lines, line, l1)) break;
|
||
injectionStats.orphans.injected++;
|
||
}
|
||
|
||
if (lines.length) {
|
||
sections.push(`[记忆碎片]\n${lines.join("\n")}`);
|
||
total.used += l1.used;
|
||
injectionStats.orphans.tokens = l1.used;
|
||
}
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
// [人物弧光]：只保留 USER + queryEntities
|
||
// ─────────────────────────────────────────────────────────────────────
|
||
if (data.arcs?.length && total.used < total.max) {
|
||
const { name1 } = getContext();
|
||
const userName = String(name1 || "").trim();
|
||
|
||
const relevantEntities = new Set(
|
||
[userName, ...(queryEntities || [])]
|
||
.map(s => String(s || "").trim())
|
||
.filter(Boolean)
|
||
);
|
||
|
||
const filteredArcs = (data.arcs || []).filter(a => {
|
||
const arcName = String(a?.name || "").trim();
|
||
return arcName && relevantEntities.has(arcName);
|
||
});
|
||
|
||
if (filteredArcs.length) {
|
||
const arcLines = filteredArcs.map(formatArcLine);
|
||
const arcText = `[人物弧光]\n${arcLines.join("\n")}`;
|
||
const arcTokens = estimateTokens(arcText);
|
||
|
||
if (total.used + arcTokens <= total.max) {
|
||
sections.push(arcText);
|
||
total.used += arcTokens;
|
||
injectionStats.arcs.count = filteredArcs.length;
|
||
injectionStats.arcs.tokens = arcTokens;
|
||
}
|
||
}
|
||
}
|
||
|
||
// 组装
|
||
if (!sections.length) {
|
||
injectionStats.budget.used = total.used;
|
||
return { promptText: "", injectionLogText: "", injectionStats };
|
||
}
|
||
|
||
injectionStats.budget.used = total.used;
|
||
const promptText = `<剧情记忆>\n\n${sections.join("\n\n")}\n\n</剧情记忆>`;
|
||
const injectionLogText = formatInjectionLog(injectionStats);
|
||
|
||
return { promptText, injectionLogText, injectionStats };
|
||
}
|
||
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
// Exported API
|
||
// ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
export async function formatPromptWithMemory(store, recallResult, causalById, queryEntities = []) {
|
||
const vectorCfg = getVectorConfig();
|
||
return vectorCfg?.enabled
|
||
? await buildMemoryPromptVectorEnabled(store, recallResult, causalById, queryEntities)
|
||
: buildMemoryPromptVectorDisabled(store);
|
||
}
|
||
|
||
export async function recallAndInjectPrompt(excludeLastAi = false, postToFrame = null) {
|
||
if (!getSettings().storySummary?.enabled) {
|
||
delete extension_prompts[SUMMARY_PROMPT_KEY];
|
||
return;
|
||
}
|
||
|
||
const { chat } = getContext();
|
||
const store = getSummaryStore();
|
||
|
||
if (!store?.json) {
|
||
delete extension_prompts[SUMMARY_PROMPT_KEY];
|
||
return;
|
||
}
|
||
|
||
const allEvents = store.json.events || [];
|
||
const lastIdx = store.lastSummarizedMesId ?? 0;
|
||
const length = chat?.length || 0;
|
||
|
||
if (lastIdx >= length) {
|
||
delete extension_prompts[SUMMARY_PROMPT_KEY];
|
||
return;
|
||
}
|
||
|
||
const vectorCfg = getVectorConfig();
|
||
let recallResult = { events: [], chunks: [], causalEvents: [], queryEntities: [] };
|
||
let causalById = new Map();
|
||
|
||
if (vectorCfg?.enabled) {
|
||
try {
|
||
const queryText = buildQueryText(chat, 2, excludeLastAi);
|
||
recallResult = await recallMemory(queryText, allEvents, vectorCfg, { excludeLastAi });
|
||
|
||
// Attach evidence chunks for causal events
|
||
const causalEvents = recallResult.causalEvents || [];
|
||
if (causalEvents.length > 0) {
|
||
const { chatId } = getContext();
|
||
if (chatId) {
|
||
try {
|
||
const floors = new Set();
|
||
for (const c of causalEvents) {
|
||
const r = parseFloorRange(c.event?.summary);
|
||
if (!r) continue;
|
||
for (let f = r.start; f <= r.end; f++) floors.add(f);
|
||
}
|
||
|
||
const [chunks, chunkVecs, eventVecs] = await Promise.all([
|
||
getChunksByFloors(chatId, Array.from(floors)),
|
||
getAllChunkVectors(chatId),
|
||
getAllEventVectors(chatId),
|
||
]);
|
||
|
||
const chunksMap = new Map(chunks.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 attachEvidenceToaCausalEvents(causalEvents, eventVectorMap, chunkVectorMap, chunksMap);
|
||
} catch (e) {
|
||
xbLog.warn(MODULE_ID, "Causal evidence attachment failed", e);
|
||
}
|
||
}
|
||
}
|
||
|
||
causalById = new Map(
|
||
(recallResult.causalEvents || [])
|
||
.map(c => [c?.event?.id, c])
|
||
.filter(x => x[0])
|
||
);
|
||
} catch (e) {
|
||
xbLog.error(MODULE_ID, "召回失败", e);
|
||
}
|
||
}
|
||
|
||
const result = await injectPrompt(
|
||
store,
|
||
recallResult,
|
||
chat,
|
||
causalById,
|
||
recallResult?.queryEntities || []
|
||
);
|
||
|
||
if (postToFrame) {
|
||
const recallLog = recallResult.logText || "";
|
||
const injectionLog = result?.injectionLogText || "";
|
||
postToFrame({ type: "RECALL_LOG", text: recallLog + injectionLog });
|
||
}
|
||
}
|
||
|
||
export function updateSummaryExtensionPrompt() {
|
||
if (!getSettings().storySummary?.enabled) {
|
||
delete extension_prompts[SUMMARY_PROMPT_KEY];
|
||
return;
|
||
}
|
||
|
||
const { chat } = getContext();
|
||
const store = getSummaryStore();
|
||
|
||
if (!store?.json || (store.lastSummarizedMesId ?? 0) >= (chat?.length || 0)) {
|
||
delete extension_prompts[SUMMARY_PROMPT_KEY];
|
||
return;
|
||
}
|
||
|
||
// 注意：这里保持“快速注入”以降低频繁触发时的开销（不做预算装配/DB批量拉取）
|
||
// 真正的预算驱动装配在 recallAndInjectPrompt() 中执行。
|
||
injectPrompt(store, { events: [], chunks: [], causalEvents: [], queryEntities: [] }, chat, new Map(), []);
|
||
}
|
||
|
||
async function injectPrompt(store, recallResult, chat, causalById, queryEntities = []) {
|
||
const length = chat?.length || 0;
|
||
|
||
const result = await formatPromptWithMemory(store, recallResult, causalById, queryEntities);
|
||
let text = result?.promptText || "";
|
||
const injectionLogText = result?.injectionLogText || "";
|
||
|
||
const cfg = getSummaryPanelConfig();
|
||
if (cfg.trigger?.wrapperHead) text = cfg.trigger.wrapperHead + "\n" + text;
|
||
if (cfg.trigger?.wrapperTail) text = text + "\n" + cfg.trigger.wrapperTail;
|
||
|
||
if (!text.trim()) {
|
||
delete extension_prompts[SUMMARY_PROMPT_KEY];
|
||
return { injectionLogText: "" };
|
||
}
|
||
|
||
const lastIdx = store.lastSummarizedMesId ?? 0;
|
||
let depth = length - lastIdx - 1;
|
||
if (depth < 0) depth = 0;
|
||
|
||
if (cfg.trigger?.forceInsertAtEnd) depth = 10000;
|
||
|
||
extension_prompts[SUMMARY_PROMPT_KEY] = {
|
||
value: text,
|
||
position: extension_prompt_types.IN_CHAT,
|
||
depth,
|
||
role: extension_prompt_roles.SYSTEM,
|
||
};
|
||
|
||
return { injectionLogText };
|
||
}
|
||
|
||
export function clearSummaryExtensionPrompt() {
|
||
delete extension_prompts[SUMMARY_PROMPT_KEY];
|
||
}
|