fix: qwen thinking toggle and recall log styles
This commit is contained in:
@@ -1,5 +1,5 @@
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
// Story Summary - Prompt Injection (v2 - DSL 版)
|
||||
// Story Summary - Prompt Injection (v3 - DSL 版 + Orphan 分组修复)
|
||||
// - 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
|
||||
// ═══════════════════════════════════════════════════════════════════════════
|
||||
@@ -23,10 +23,6 @@ const MODULE_ID = "summaryPrompt";
|
||||
let lastRecallFailAt = 0;
|
||||
const RECALL_FAIL_COOLDOWN_MS = 10_000;
|
||||
|
||||
/**
|
||||
* 检查是否可以通知召回失败
|
||||
* @returns {boolean}
|
||||
*/
|
||||
function canNotifyRecallFail() {
|
||||
const now = Date.now();
|
||||
if (now - lastRecallFailAt < RECALL_FAIL_COOLDOWN_MS) return false;
|
||||
@@ -50,11 +46,6 @@ const TOP_N_STAR = 5;
|
||||
// 工具函数
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 估算 token 数量
|
||||
* @param {string} text - 文本
|
||||
* @returns {number} token 数
|
||||
*/
|
||||
function estimateTokens(text) {
|
||||
if (!text) return 0;
|
||||
const s = String(text);
|
||||
@@ -62,13 +53,6 @@ function estimateTokens(text) {
|
||||
return Math.ceil(zh + (s.length - zh) / 4);
|
||||
}
|
||||
|
||||
/**
|
||||
* 带预算控制的行推入
|
||||
* @param {Array} lines - 行数组
|
||||
* @param {string} text - 文本
|
||||
* @param {object} state - 预算状态 {used, max}
|
||||
* @returns {boolean} 是否成功
|
||||
*/
|
||||
function pushWithBudget(lines, text, state) {
|
||||
const t = estimateTokens(text);
|
||||
if (state.used + t > state.max) return false;
|
||||
@@ -77,12 +61,6 @@ function pushWithBudget(lines, text, state) {
|
||||
return true;
|
||||
}
|
||||
|
||||
/**
|
||||
* 计算余弦相似度
|
||||
* @param {Array} a - 向量 a
|
||||
* @param {Array} 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;
|
||||
@@ -94,11 +72,6 @@ function cosineSimilarity(a, b) {
|
||||
return nA && nB ? dot / (Math.sqrt(nA) * Math.sqrt(nB)) : 0;
|
||||
}
|
||||
|
||||
/**
|
||||
* 解析楼层范围
|
||||
* @param {string} summary - 摘要文本
|
||||
* @returns {object|null} {start, end}
|
||||
*/
|
||||
function parseFloorRange(summary) {
|
||||
if (!summary) return null;
|
||||
const match = String(summary).match(/\(#(\d+)(?:-(\d+))?\)/);
|
||||
@@ -108,22 +81,12 @@ function parseFloorRange(summary) {
|
||||
return { start, end };
|
||||
}
|
||||
|
||||
/**
|
||||
* 清理摘要中的楼层标记
|
||||
* @param {string} summary - 摘要文本
|
||||
* @returns {string} 清理后的文本
|
||||
*/
|
||||
function cleanSummary(summary) {
|
||||
return String(summary || "")
|
||||
.replace(/\s*\(#\d+(?:-\d+)?\)\s*$/, "")
|
||||
.trim();
|
||||
}
|
||||
|
||||
/**
|
||||
* 规范化字符串（用于比较）
|
||||
* @param {string} s - 字符串
|
||||
* @returns {string} 规范化后的字符串
|
||||
*/
|
||||
function normalize(s) {
|
||||
return String(s || '')
|
||||
.normalize('NFKC')
|
||||
@@ -136,22 +99,11 @@ function normalize(s) {
|
||||
// 上下文配对工具函数
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 获取上下文楼层
|
||||
* @param {object} chunk - chunk 对象
|
||||
* @returns {number} 配对楼层，-1 表示无效
|
||||
*/
|
||||
function getContextFloor(chunk) {
|
||||
if (chunk.isL0) return -1;
|
||||
return chunk.isUser ? chunk.floor + 1 : chunk.floor - 1;
|
||||
}
|
||||
|
||||
/**
|
||||
* 选择配对 chunk
|
||||
* @param {Array} candidates - 候选 chunks
|
||||
* @param {object} mainChunk - 主 chunk
|
||||
* @returns {object|null} 配对 chunk
|
||||
*/
|
||||
function pickContextChunk(candidates, mainChunk) {
|
||||
if (!candidates?.length) return null;
|
||||
const targetIsUser = !mainChunk.isUser;
|
||||
@@ -160,12 +112,6 @@ function pickContextChunk(candidates, mainChunk) {
|
||||
return candidates[0];
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式化上下文 chunk 行
|
||||
* @param {object} chunk - chunk 对象
|
||||
* @param {boolean} isAbove - 是否在主 chunk 上方
|
||||
* @returns {string} 格式化的行
|
||||
*/
|
||||
function formatContextChunkLine(chunk, isAbove) {
|
||||
const { name1, name2 } = getContext();
|
||||
const speaker = chunk.isUser ? (name1 || "用户") : (chunk.speaker || name2 || "角色");
|
||||
@@ -178,10 +124,6 @@ function formatContextChunkLine(chunk, isAbove) {
|
||||
// 系统前导与后缀
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 构建系统前导
|
||||
* @returns {string}
|
||||
*/
|
||||
function buildSystemPreamble() {
|
||||
return [
|
||||
"以上是还留在眼前的对话",
|
||||
@@ -193,10 +135,6 @@ function buildSystemPreamble() {
|
||||
].join("\n");
|
||||
}
|
||||
|
||||
/**
|
||||
* 构建后缀
|
||||
* @returns {string}
|
||||
*/
|
||||
function buildPostscript() {
|
||||
return [
|
||||
"",
|
||||
@@ -208,28 +146,20 @@ function buildPostscript() {
|
||||
// L1 Facts 分层过滤
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 从 store 获取所有已知角色名
|
||||
* @param {object} store - summary store
|
||||
* @returns {Set<string>} 角色名集合（规范化后）
|
||||
*/
|
||||
function getKnownCharacters(store) {
|
||||
const names = new Set();
|
||||
|
||||
// 从 arcs 获取
|
||||
const arcs = store?.json?.arcs || [];
|
||||
for (const a of arcs) {
|
||||
if (a.name) names.add(normalize(a.name));
|
||||
}
|
||||
|
||||
// 从 characters.main 获取
|
||||
const main = store?.json?.characters?.main || [];
|
||||
for (const m of main) {
|
||||
const name = typeof m === 'string' ? m : m.name;
|
||||
if (name) names.add(normalize(name));
|
||||
}
|
||||
|
||||
// 从当前角色获取
|
||||
const { name1, name2 } = getContext();
|
||||
if (name1) names.add(normalize(name1));
|
||||
if (name2) names.add(normalize(name2));
|
||||
@@ -237,77 +167,42 @@ function getKnownCharacters(store) {
|
||||
return names;
|
||||
}
|
||||
|
||||
/**
|
||||
* 解析关系类 fact 的目标人物
|
||||
* @param {string} predicate - 谓词，如 "对蓝袖的看法"
|
||||
* @returns {string|null} 目标人物名
|
||||
*/
|
||||
function parseRelationTarget(predicate) {
|
||||
const match = String(predicate || '').match(/^对(.+)的/);
|
||||
return match ? match[1] : null;
|
||||
}
|
||||
|
||||
/**
|
||||
* 过滤 facts（分层策略）
|
||||
*
|
||||
* 规则：
|
||||
* - isState=true：全量保留
|
||||
* - 关系类（谓词匹配 /^对.+的/）：from 或 to 在 focus 中
|
||||
* - 人物状态类（主体是已知角色名）：主体在 focus 中
|
||||
* - 其他（物品/地点/规则）：全量保留
|
||||
*
|
||||
* @param {Array} facts - 所有 facts
|
||||
* @param {Array} focusEntities - 焦点实体
|
||||
* @param {Set} knownCharacters - 已知角色名集合
|
||||
* @returns {Array} 过滤后的 facts
|
||||
*/
|
||||
function filterFactsByRelevance(facts, focusEntities, knownCharacters) {
|
||||
if (!facts?.length) return [];
|
||||
|
||||
const focusSet = new Set((focusEntities || []).map(normalize));
|
||||
|
||||
return facts.filter(f => {
|
||||
// 1. isState=true：全量保留
|
||||
if (f._isState === true) return true;
|
||||
|
||||
// 2. 关系类：from 或 to 在 focus 中
|
||||
if (isRelationFact(f)) {
|
||||
const from = normalize(f.s);
|
||||
const target = parseRelationTarget(f.p);
|
||||
const to = target ? normalize(target) : '';
|
||||
|
||||
// 任一方在 focus 中即保留
|
||||
if (focusSet.has(from) || focusSet.has(to)) return true;
|
||||
|
||||
// 都不在 focus 中则过滤
|
||||
return false;
|
||||
}
|
||||
|
||||
// 3. 主体是已知角色名：检查是否在 focus 中
|
||||
const subjectNorm = normalize(f.s);
|
||||
if (knownCharacters.has(subjectNorm)) {
|
||||
return focusSet.has(subjectNorm);
|
||||
}
|
||||
|
||||
// 4. 主体不是人名（物品/地点/规则等）：保留
|
||||
return true;
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式化 facts 用于注入
|
||||
* @param {Array} facts - facts 数组
|
||||
* @param {Array} focusEntities - 焦点实体
|
||||
* @param {Set} knownCharacters - 已知角色名集合
|
||||
* @returns {Array} 格式化后的行
|
||||
*/
|
||||
function formatFactsForInjection(facts, focusEntities, knownCharacters) {
|
||||
// 先过滤
|
||||
const filtered = filterFactsByRelevance(facts, focusEntities, knownCharacters);
|
||||
|
||||
if (!filtered.length) return [];
|
||||
|
||||
// 按 since 降序排序（最新的优先）
|
||||
return filtered
|
||||
.sort((a, b) => (b.since || 0) - (a.since || 0))
|
||||
.map(f => {
|
||||
@@ -323,11 +218,6 @@ function formatFactsForInjection(facts, focusEntities, knownCharacters) {
|
||||
// 格式化函数
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 格式化角色弧光行
|
||||
* @param {object} a - 弧光对象
|
||||
* @returns {string}
|
||||
*/
|
||||
function formatArcLine(a) {
|
||||
const moments = (a.moments || [])
|
||||
.map(m => (typeof m === "string" ? m : m.text))
|
||||
@@ -339,11 +229,6 @@ function formatArcLine(a) {
|
||||
return `- ${a.name}：${a.trajectory}`;
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式化 chunk 完整行
|
||||
* @param {object} c - chunk 对象
|
||||
* @returns {string}
|
||||
*/
|
||||
function formatChunkFullLine(c) {
|
||||
const { name1, name2 } = getContext();
|
||||
|
||||
@@ -355,38 +240,6 @@ function formatChunkFullLine(c) {
|
||||
return `› #${c.floor + 1} [${speaker}] ${String(c.text || "").trim()}`;
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式化带上下文的 chunk
|
||||
* @param {object} mainChunk - 主 chunk
|
||||
* @param {object|null} contextChunk - 上下文 chunk
|
||||
* @returns {Array} 格式化的行数组
|
||||
*/
|
||||
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;
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式化因果事件行
|
||||
* @param {object} causalItem - 因果项
|
||||
* @param {Map} causalById - 因果映射
|
||||
* @returns {string}
|
||||
*/
|
||||
function formatCausalEventLine(causalItem, causalById) {
|
||||
const ev = causalItem?.event || {};
|
||||
const depth = Math.max(1, Math.min(9, causalItem?._causalDepth || 1));
|
||||
@@ -415,22 +268,11 @@ function formatCausalEventLine(causalItem, causalById) {
|
||||
return lines.join("\n");
|
||||
}
|
||||
|
||||
/**
|
||||
* 重新编号事件文本
|
||||
* @param {string} text - 事件文本
|
||||
* @param {number} newIndex - 新编号
|
||||
* @returns {string}
|
||||
*/
|
||||
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} ev - 事件对象
|
||||
* @returns {number}
|
||||
*/
|
||||
function getEventSortKey(ev) {
|
||||
const r = parseFloorRange(ev?.summary);
|
||||
if (r) return r.start;
|
||||
@@ -438,20 +280,98 @@ function getEventSortKey(ev) {
|
||||
return m ? parseInt(m[1], 10) : Number.MAX_SAFE_INTEGER;
|
||||
}
|
||||
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
// 按楼层分组装配 orphan chunks（修复上下文重复）
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
function assembleOrphansByFloor(orphanCandidates, contextChunksByFloor, budget) {
|
||||
if (!orphanCandidates?.length) {
|
||||
return { lines: [], l0Count: 0, contextPairsCount: 0 };
|
||||
}
|
||||
|
||||
// 1. 按楼层分组
|
||||
const byFloor = new Map();
|
||||
for (const c of orphanCandidates) {
|
||||
const arr = byFloor.get(c.floor) || [];
|
||||
arr.push(c);
|
||||
byFloor.set(c.floor, arr);
|
||||
}
|
||||
|
||||
// 2. 楼层内按 chunkIdx 排序
|
||||
for (const [, chunks] of byFloor) {
|
||||
chunks.sort((a, b) => (a.chunkIdx ?? 0) - (b.chunkIdx ?? 0));
|
||||
}
|
||||
|
||||
// 3. 按楼层顺序装配
|
||||
const floorsSorted = Array.from(byFloor.keys()).sort((a, b) => a - b);
|
||||
|
||||
const lines = [];
|
||||
let l0Count = 0;
|
||||
let contextPairsCount = 0;
|
||||
|
||||
for (const floor of floorsSorted) {
|
||||
const chunks = byFloor.get(floor);
|
||||
if (!chunks?.length) continue;
|
||||
|
||||
// 分离 L0 和 L1
|
||||
const l0Chunks = chunks.filter(c => c.isL0);
|
||||
const l1Chunks = chunks.filter(c => !c.isL0);
|
||||
|
||||
// L0 直接输出（不需要上下文）
|
||||
for (const c of l0Chunks) {
|
||||
const line = formatChunkFullLine(c);
|
||||
if (!pushWithBudget(lines, line, budget)) {
|
||||
return { lines, l0Count, contextPairsCount };
|
||||
}
|
||||
l0Count++;
|
||||
}
|
||||
|
||||
// L1 按楼层统一处理
|
||||
if (l1Chunks.length > 0) {
|
||||
const firstChunk = l1Chunks[0];
|
||||
const pairFloor = getContextFloor(firstChunk);
|
||||
const pairCandidates = contextChunksByFloor.get(pairFloor) || [];
|
||||
const contextChunk = pickContextChunk(pairCandidates, firstChunk);
|
||||
|
||||
// 上下文在前
|
||||
if (contextChunk && contextChunk.floor < floor) {
|
||||
const contextLine = formatContextChunkLine(contextChunk, true);
|
||||
if (!pushWithBudget(lines, contextLine, budget)) {
|
||||
return { lines, l0Count, contextPairsCount };
|
||||
}
|
||||
contextPairsCount++;
|
||||
}
|
||||
|
||||
// 输出该楼层所有 L1 chunks
|
||||
for (const c of l1Chunks) {
|
||||
const line = formatChunkFullLine(c);
|
||||
if (!pushWithBudget(lines, line, budget)) {
|
||||
return { lines, l0Count, contextPairsCount };
|
||||
}
|
||||
}
|
||||
|
||||
// 上下文在后
|
||||
if (contextChunk && contextChunk.floor > floor) {
|
||||
const contextLine = formatContextChunkLine(contextChunk, false);
|
||||
if (!pushWithBudget(lines, contextLine, budget)) {
|
||||
return { lines, l0Count, contextPairsCount };
|
||||
}
|
||||
contextPairsCount++;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return { lines, l0Count, contextPairsCount };
|
||||
}
|
||||
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
// 非向量模式
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 构建非向量模式的 prompt
|
||||
* @param {object} store - summary store
|
||||
* @returns {string}
|
||||
*/
|
||||
function buildNonVectorPrompt(store) {
|
||||
const data = store.json || {};
|
||||
const sections = [];
|
||||
|
||||
// L1 facts（非向量模式不做分层过滤，全量注入）
|
||||
const allFacts = getFacts();
|
||||
const factLines = allFacts
|
||||
.filter(f => !f.retracted)
|
||||
@@ -494,10 +414,6 @@ function buildNonVectorPrompt(store) {
|
||||
);
|
||||
}
|
||||
|
||||
/**
|
||||
* 构建非向量模式的注入文本
|
||||
* @returns {string}
|
||||
*/
|
||||
export function buildNonVectorPromptText() {
|
||||
if (!getSettings().storySummary?.enabled) {
|
||||
return "";
|
||||
@@ -524,16 +440,6 @@ export function buildNonVectorPromptText() {
|
||||
// 向量模式：预算装配
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 构建向量模式的 prompt
|
||||
* @param {object} store - summary store
|
||||
* @param {object} recallResult - 召回结果
|
||||
* @param {Map} causalById - 因果映射
|
||||
* @param {Array} focusEntities - 焦点实体
|
||||
* @param {object} meta - 元数据
|
||||
* @param {object} metrics - 指标对象
|
||||
* @returns {Promise<object>} {promptText, injectionLogText, injectionStats, metrics}
|
||||
*/
|
||||
async function buildVectorPrompt(store, recallResult, causalById, focusEntities = [], meta = null, metrics = null) {
|
||||
const T_Start = performance.now();
|
||||
|
||||
@@ -541,7 +447,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
const data = store.json || {};
|
||||
const total = { used: 0, max: MAIN_BUDGET_MAX };
|
||||
|
||||
// 预装配容器
|
||||
const assembled = {
|
||||
facts: { lines: [], tokens: 0 },
|
||||
arcs: { lines: [], tokens: 0 },
|
||||
@@ -573,7 +478,7 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
};
|
||||
|
||||
// ═══════════════════════════════════════════════════════════════════════
|
||||
// [优先级 1] 世界约束 - 最高优先级（带分层过滤）
|
||||
// [优先级 1] 世界约束
|
||||
// ═══════════════════════════════════════════════════════════════════════
|
||||
|
||||
const T_L1_Start = performance.now();
|
||||
@@ -582,7 +487,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
const knownCharacters = getKnownCharacters(store);
|
||||
const factLines = formatFactsForInjection(allFacts, focusEntities, knownCharacters);
|
||||
|
||||
// METRICS: L1 指标
|
||||
if (metrics) {
|
||||
metrics.l1.factsTotal = allFacts.length;
|
||||
metrics.l1.factsFiltered = allFacts.length - factLines.length;
|
||||
@@ -599,7 +503,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
injectionStats.facts.tokens = l1Budget.used;
|
||||
injectionStats.facts.filtered = allFacts.length - factLines.length;
|
||||
|
||||
// METRICS
|
||||
if (metrics) {
|
||||
metrics.l1.factsInjected = assembled.facts.lines.length;
|
||||
metrics.l1.tokens = l1Budget.used;
|
||||
@@ -613,7 +516,7 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
}
|
||||
|
||||
// ═══════════════════════════════════════════════════════════════════════
|
||||
// [优先级 2] 人物弧光 - 预留预算
|
||||
// [优先级 2] 人物弧光
|
||||
// ═══════════════════════════════════════════════════════════════════════
|
||||
|
||||
if (data.arcs?.length && total.used < total.max) {
|
||||
@@ -652,13 +555,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
const chunks = recallResult?.chunks || [];
|
||||
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
|
||||
*/
|
||||
|
||||
// 优先 L0 虚拟 chunk，否则按 chunkIdx 选第一个
|
||||
function pickBestChunkForEvent(eventObj) {
|
||||
const range = parseFloorRange(eventObj?.summary);
|
||||
if (!range) return null;
|
||||
@@ -667,27 +563,18 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
for (const c of chunks) {
|
||||
if (usedChunkIds.has(c.chunkId)) continue;
|
||||
if (c.floor < range.start || c.floor > range.end) continue;
|
||||
|
||||
|
||||
if (!best) {
|
||||
best = c;
|
||||
} else if (c.isL0 && !best.isL0) {
|
||||
// L0 优先
|
||||
best = c;
|
||||
} else if (c.isL0 === best.isL0 && (c.chunkIdx ?? 0) < (best.chunkIdx ?? 0)) {
|
||||
// 同类型按 chunkIdx 选靠前的
|
||||
best = c;
|
||||
}
|
||||
}
|
||||
return best;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式化带证据的事件
|
||||
* @param {object} e - 事件召回项
|
||||
* @param {number} idx - 索引
|
||||
* @param {object|null} chunk - 证据 chunk
|
||||
* @returns {string}
|
||||
*/
|
||||
function formatEventWithEvidence(e, idx, chunk) {
|
||||
const ev = e.event || {};
|
||||
const time = ev.timeLabel || "";
|
||||
@@ -775,7 +662,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
});
|
||||
}
|
||||
|
||||
// 重排
|
||||
selectedDirect.sort((a, b) => getEventSortKey(a.event) - getEventSortKey(b.event));
|
||||
selectedSimilar.sort((a, b) => getEventSortKey(a.event) - getEventSortKey(b.event));
|
||||
|
||||
@@ -829,47 +715,22 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
}
|
||||
|
||||
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: Math.min(ORPHAN_MAX, total.max - total.used) };
|
||||
let l0Count = 0;
|
||||
let contextPairsCount = 0;
|
||||
|
||||
for (const c of orphans) {
|
||||
if (c.isL0) {
|
||||
const line = formatChunkFullLine(c);
|
||||
if (!pushWithBudget(assembled.orphans.lines, line, l1Budget)) break;
|
||||
injectionStats.orphans.injected++;
|
||||
l0Count++;
|
||||
continue;
|
||||
}
|
||||
|
||||
const pairFloor = getContextFloor(c);
|
||||
const pairCandidates = contextChunksByFloor.get(pairFloor) || [];
|
||||
const contextChunk = pickContextChunk(pairCandidates, c);
|
||||
|
||||
const formattedLines = formatChunkWithContext(c, contextChunk);
|
||||
|
||||
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++;
|
||||
}
|
||||
const result = assembleOrphansByFloor(
|
||||
orphanCandidates.sort((a, b) => (a.floor - b.floor) || ((a.chunkIdx ?? 0) - (b.chunkIdx ?? 0))),
|
||||
contextChunksByFloor,
|
||||
l1Budget
|
||||
);
|
||||
|
||||
assembled.orphans.lines = result.lines;
|
||||
assembled.orphans.tokens = l1Budget.used;
|
||||
total.used += l1Budget.used;
|
||||
|
||||
injectionStats.orphans.injected = result.lines.length;
|
||||
injectionStats.orphans.tokens = l1Budget.used;
|
||||
injectionStats.orphans.l0Count = l0Count;
|
||||
injectionStats.orphans.contextPairs = contextPairsCount;
|
||||
injectionStats.orphans.l0Count = result.l0Count;
|
||||
injectionStats.orphans.contextPairs = result.contextPairsCount;
|
||||
}
|
||||
|
||||
// ═══════════════════════════════════════════════════════════════════════
|
||||
@@ -891,7 +752,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
if (pairFloor >= 0) recentContextFloors.add(pairFloor);
|
||||
}
|
||||
|
||||
let recentContextChunksByFloor = new Map();
|
||||
if (chatId && recentContextFloors.size > 0) {
|
||||
const newFloors = Array.from(recentContextFloors).filter(f => !contextChunksByFloor.has(f));
|
||||
if (newFloors.length > 0) {
|
||||
@@ -907,47 +767,25 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
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)));
|
||||
if (recentOrphanCandidates.length) {
|
||||
const recentBudget = { used: 0, max: RECENT_ORPHAN_MAX };
|
||||
|
||||
const recentBudget = { used: 0, max: RECENT_ORPHAN_MAX };
|
||||
let recentContextPairsCount = 0;
|
||||
const result = assembleOrphansByFloor(
|
||||
recentOrphanCandidates.sort((a, b) => (a.floor - b.floor) || ((a.chunkIdx ?? 0) - (b.chunkIdx ?? 0))),
|
||||
contextChunksByFloor,
|
||||
recentBudget
|
||||
);
|
||||
|
||||
for (const c of recentOrphans) {
|
||||
if (c.isL0) {
|
||||
const line = formatChunkFullLine(c);
|
||||
if (!pushWithBudget(assembled.recentOrphans.lines, line, recentBudget)) break;
|
||||
recentOrphanStats.injected++;
|
||||
continue;
|
||||
}
|
||||
assembled.recentOrphans.lines = result.lines;
|
||||
assembled.recentOrphans.tokens = recentBudget.used;
|
||||
|
||||
const pairFloor = getContextFloor(c);
|
||||
const pairCandidates = recentContextChunksByFloor.get(pairFloor) || [];
|
||||
const contextChunk = pickContextChunk(pairCandidates, c);
|
||||
|
||||
const formattedLines = formatChunkWithContext(c, contextChunk);
|
||||
|
||||
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++;
|
||||
recentOrphanStats.injected = result.lines.length;
|
||||
recentOrphanStats.tokens = recentBudget.used;
|
||||
recentOrphanStats.floorRange = `${recentStart + 1}~${recentEnd + 1}楼`;
|
||||
recentOrphanStats.contextPairs = result.contextPairsCount;
|
||||
}
|
||||
|
||||
assembled.recentOrphans.tokens = recentBudget.used;
|
||||
recentOrphanStats.tokens = recentBudget.used;
|
||||
recentOrphanStats.floorRange = `${recentStart + 1}~${recentEnd + 1}楼`;
|
||||
recentOrphanStats.contextPairs = recentContextPairsCount;
|
||||
}
|
||||
|
||||
// ═══════════════════════════════════════════════════════════════════════
|
||||
@@ -990,9 +828,7 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
`<剧情记忆>\n\n${sections.join("\n\n")}\n\n</剧情记忆>\n` +
|
||||
`${buildPostscript()}`;
|
||||
|
||||
// METRICS: 更新 L4 和 Budget 指标
|
||||
if (metrics) {
|
||||
// L4 指标
|
||||
metrics.l4.sectionsIncluded = [];
|
||||
if (assembled.facts.lines.length) metrics.l4.sectionsIncluded.push('constraints');
|
||||
if (assembled.events.direct.length) metrics.l4.sectionsIncluded.push('direct_events');
|
||||
@@ -1004,7 +840,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
metrics.l4.formattingTime = Math.round(performance.now() - T_L4_Start);
|
||||
metrics.timing.l4Formatting = metrics.l4.formattingTime;
|
||||
|
||||
// Budget 指标
|
||||
metrics.budget.total = total.used + (assembled.recentOrphans.tokens || 0);
|
||||
metrics.budget.limit = TOTAL_BUDGET_MAX;
|
||||
metrics.budget.utilization = Math.round(metrics.budget.total / TOTAL_BUDGET_MAX * 100);
|
||||
@@ -1016,13 +851,11 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
arcs: assembled.arcs.tokens,
|
||||
};
|
||||
|
||||
// L3 额外指标
|
||||
metrics.l3.tokens = injectionStats.orphans.tokens + (recentOrphanStats.tokens || 0);
|
||||
metrics.l3.contextPairsAdded = injectionStats.orphans.contextPairs + recentOrphanStats.contextPairs;
|
||||
metrics.l3.assemblyTime = Math.round(performance.now() - T_Start - (metrics.timing.l1Constraints || 0) - metrics.l4.formattingTime);
|
||||
metrics.timing.l3Assembly = metrics.l3.assemblyTime;
|
||||
|
||||
// 质量指标
|
||||
const totalFacts = allFacts.length;
|
||||
metrics.quality.constraintCoverage = totalFacts > 0
|
||||
? Math.round(assembled.facts.lines.length / totalFacts * 100)
|
||||
@@ -1035,7 +868,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
? Math.round(chunksWithEvents / totalChunks * 100)
|
||||
: 0;
|
||||
|
||||
// 检测问题
|
||||
metrics.quality.potentialIssues = detectIssues(metrics);
|
||||
}
|
||||
|
||||
@@ -1046,13 +878,6 @@ async function buildVectorPrompt(store, recallResult, causalById, focusEntities
|
||||
// 因果证据补充
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 为因果事件附加证据 chunk
|
||||
* @param {Array} causalEvents - 因果事件列表
|
||||
* @param {Map} eventVectorMap - 事件向量映射
|
||||
* @param {Map} chunkVectorMap - chunk 向量映射
|
||||
* @param {Map} chunksMap - chunk 映射
|
||||
*/
|
||||
async function attachEvidenceToCausalEvents(causalEvents, eventVectorMap, chunkVectorMap, chunksMap) {
|
||||
for (const c of causalEvents) {
|
||||
c._evidenceChunk = null;
|
||||
@@ -1100,12 +925,6 @@ async function attachEvidenceToCausalEvents(causalEvents, eventVectorMap, chunkV
|
||||
// 向量模式：召回 + 注入
|
||||
// ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
/**
|
||||
* 构建向量模式的注入文本
|
||||
* @param {boolean} excludeLastAi - 是否排除最后一条 AI 消息
|
||||
* @param {object} hooks - 钩子 {postToFrame, echo, pendingUserMessage}
|
||||
* @returns {Promise<object>} {text, logText}
|
||||
*/
|
||||
export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
|
||||
const { postToFrame = null, echo = null, pendingUserMessage = null } = hooks;
|
||||
|
||||
@@ -1156,7 +975,6 @@ export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
|
||||
metrics: recallResult?.metrics || null,
|
||||
};
|
||||
|
||||
// 给因果事件挂证据
|
||||
const causalEvents = recallResult.causalEvents || [];
|
||||
if (causalEvents.length > 0) {
|
||||
if (chatId) {
|
||||
@@ -1228,7 +1046,6 @@ export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
|
||||
return { text: "", logText: "\n[Vector Recall Empty]\nNo recall candidates / vectors not ready.\n" };
|
||||
}
|
||||
|
||||
// 拼装向量 prompt，传入 focusEntities 和 metrics
|
||||
const { promptText, metrics: promptMetrics } = await buildVectorPrompt(
|
||||
store,
|
||||
recallResult,
|
||||
@@ -1238,16 +1055,13 @@ export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
|
||||
recallResult?.metrics || null
|
||||
);
|
||||
|
||||
// wrapper
|
||||
const cfg = getSummaryPanelConfig();
|
||||
let finalText = String(promptText || "");
|
||||
if (cfg.trigger?.wrapperHead) finalText = cfg.trigger.wrapperHead + "\n" + finalText;
|
||||
if (cfg.trigger?.wrapperTail) finalText = finalText + "\n" + cfg.trigger.wrapperTail;
|
||||
|
||||
// METRICS: 生成完整的指标日志
|
||||
const metricsLogText = promptMetrics ? formatMetricsLog(promptMetrics) : '';
|
||||
|
||||
// 发给 iframe
|
||||
if (postToFrame) {
|
||||
postToFrame({ type: "RECALL_LOG", text: metricsLogText });
|
||||
}
|
||||
|
||||
@@ -1455,23 +1455,25 @@ h1 span {
|
||||
}
|
||||
|
||||
#recall-log-content {
|
||||
flex: 1;
|
||||
min-height: 0;
|
||||
white-space: pre-wrap;
|
||||
font-family: 'SF Mono', Monaco, Consolas, 'Courier New', monospace;
|
||||
font-family: 'Consolas', 'Monaco', 'SF Mono', monospace;
|
||||
font-size: 12px;
|
||||
line-height: 1.6;
|
||||
background: var(--bg3);
|
||||
padding: 16px;
|
||||
border-radius: 4px;
|
||||
overflow-y: auto;
|
||||
color: #e8e8e8;
|
||||
white-space: pre-wrap !important;
|
||||
overflow-x: hidden !important;
|
||||
word-break: break-word;
|
||||
overflow-wrap: break-word;
|
||||
-webkit-font-smoothing: antialiased;
|
||||
-moz-osx-font-smoothing: grayscale;
|
||||
}
|
||||
|
||||
.recall-empty {
|
||||
color: var(--txt3);
|
||||
color: #999;
|
||||
text-align: center;
|
||||
padding: 40px;
|
||||
font-style: italic;
|
||||
font-size: .8125rem;
|
||||
line-height: 1.8;
|
||||
}
|
||||
|
||||
/* 移动端适配 */
|
||||
@@ -1483,9 +1485,11 @@ h1 span {
|
||||
border-radius: 0;
|
||||
}
|
||||
|
||||
.debug-log-viewer,
|
||||
#recall-log-content {
|
||||
font-size: 11px;
|
||||
padding: 12px;
|
||||
line-height: 1.5;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -2732,14 +2736,18 @@ h1 span {
|
||||
margin-bottom: 4px;
|
||||
}
|
||||
|
||||
/* ═══════════════════════════════════════════════════════════════════════════
|
||||
Recall Log / Debug Log
|
||||
═══════════════════════════════════════════════════════════════════════════ */
|
||||
|
||||
.debug-log-viewer {
|
||||
background: #1e1e1e;
|
||||
color: #d4d4d4;
|
||||
background: #1a1a1a;
|
||||
color: #e0e0e0;
|
||||
padding: 16px;
|
||||
border-radius: 8px;
|
||||
font-family: 'Consolas', 'Monaco', monospace;
|
||||
font-family: 'Consolas', 'Monaco', 'SF Mono', monospace;
|
||||
font-size: 12px;
|
||||
line-height: 1.5;
|
||||
line-height: 1.6;
|
||||
max-height: 60vh;
|
||||
overflow-y: auto;
|
||||
overflow-x: hidden;
|
||||
@@ -2749,7 +2757,7 @@ h1 span {
|
||||
}
|
||||
|
||||
.recall-empty {
|
||||
color: var(--txt3);
|
||||
color: #999;
|
||||
text-align: center;
|
||||
padding: 40px;
|
||||
font-style: italic;
|
||||
@@ -2884,15 +2892,6 @@ h1 span {
|
||||
Metrics Log Styling
|
||||
═══════════════════════════════════════════════════════════════════════════ */
|
||||
|
||||
#recall-log-content {
|
||||
font-family: 'SF Mono', Monaco, Consolas, 'Courier New', monospace;
|
||||
font-size: 11px;
|
||||
line-height: 1.5;
|
||||
white-space: pre;
|
||||
overflow-x: auto;
|
||||
tab-size: 4;
|
||||
}
|
||||
|
||||
#recall-log-content .metric-warn {
|
||||
color: #f59e0b;
|
||||
}
|
||||
|
||||
@@ -29,7 +29,7 @@ function b64UrlEncode(str) {
|
||||
|
||||
/**
|
||||
* 统一LLM调用 - 走酒馆后端（非流式）
|
||||
* 修复：assistant prefill 用 bottomassistant 参数传递
|
||||
* assistant prefill 用 bottomassistant 参数传递
|
||||
*/
|
||||
export async function callLLM(messages, options = {}) {
|
||||
const {
|
||||
@@ -46,10 +46,10 @@ export async function callLLM(messages, options = {}) {
|
||||
throw new Error('L0 requires siliconflow API key');
|
||||
}
|
||||
|
||||
// ★ 关键修复：分离 assistant prefill
|
||||
// 分离 assistant prefill
|
||||
let topMessages = [...messages];
|
||||
let assistantPrefill = '';
|
||||
|
||||
|
||||
if (topMessages.length > 0 && topMessages[topMessages.length - 1]?.role === 'assistant') {
|
||||
const lastMsg = topMessages.pop();
|
||||
assistantPrefill = lastMsg.content || '';
|
||||
@@ -70,6 +70,10 @@ export async function callLLM(messages, options = {}) {
|
||||
apipassword: apiKey,
|
||||
model: DEFAULT_L0_MODEL,
|
||||
};
|
||||
const isQwen3 = String(DEFAULT_L0_MODEL || '').includes('Qwen3');
|
||||
if (isQwen3) {
|
||||
args.enable_thinking = 'false';
|
||||
}
|
||||
|
||||
// ★ 用 bottomassistant 参数传递 prefill
|
||||
if (assistantPrefill) {
|
||||
|
||||
Reference in New Issue
Block a user