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

feat(summary): update prompt display, metrics lexical gate, and edge sanitization

Browse Source
This commit is contained in:
bielie
2026-02-11 22:01:02 +08:00
parent ca117b334f
commit 9f279d902f
4 changed files with 275 additions and 306 deletions
Download Patch File Download Diff File Expand all files Collapse all files

305
modules/story-summary/vector/llm/atom-extraction.js
View File

@@ -1,5 +1,13 @@
// ============================================================================
// atom-extraction.js - L0 叙事锚点提取（三层 themes 版）
// atom-extraction.js - L0 场景锚点提取（v2 - 场景摘要 + 图结构）
//
// 设计依据：
// - BGE-M3 (BAAI, 2024): 自然语言段落检索精度最高 → semantic = 纯自然语言
// - Interpersonal Circumplex (Kiesler, 1983): 权力轴+情感轴 → dynamics 枚举
// - Labov Narrative Structure (1972): 叙事功能轴 → dynamics 枚举补充
// - TransE (Bordes, 2013): s/t/r 三元组方向性 → edges 格式
//
// 每楼层 1-2 个场景锚点（非碎片原子），60-100 字场景摘要
// ============================================================================
import { callLLM, parseJson } from './llm-service.js';
@@ -25,10 +33,30 @@ export function isBatchCancelled() {
}
// ============================================================================
// L0 提取 Prompt（三层 themes）
// dynamics 封闭枚举（8 个标签，两轴四象限 + 叙事轴）
// ============================================================================
const SYSTEM_PROMPT = `你是叙事锚点提取器。从一轮对话中提取4-8个关键锚点，用于后续语义检索。
const VALID_DYNAMICS = new Set([
// 权力轴 (Interpersonal Circumplex: Dominance-Submission)
'支配', // 控制、命令、审视、威慑、主导
'让渡', // 顺从、服从、屈服、被动、配合
// 情感轴 (Interpersonal Circumplex: Hostility-Friendliness)
'亲密', // 温柔、关怀、依赖、信任、连接
'敌意', // 对抗、拒绝、攻击、嘲讽、排斥
// 叙事轴 (Labov Narrative Structure)
'揭示', // 真相、发现、告白、暴露、秘密
'决意', // 选择、承诺、放弃、宣言、转折
'张力', // 悬念、对峙、暗涌、不安、等待
'丧失', // 分离、死亡、破碎、遗憾、崩塌
]);
// ============================================================================
// L0 提取 Prompt
// ============================================================================
const SYSTEM_PROMPT = `你是场景摘要器。从一轮对话中提取1-2个场景锚点，用于语义检索和关系追踪。
输入格式：
<round>
@@ -37,126 +65,50 @@ const SYSTEM_PROMPT = `你是叙事锚点提取器。从一轮对话中提取4-8
</round>
只输出严格JSON：
{"atoms":[{"t":"类型","s":"主体","o":"客体","v":"谓词","l":"地点","f":"来源","th":{"fn":[],"pt":[],"kw":[]}}]}
## 类型（t）
- emo: 情绪状态变化
- act: 关键动作/行为
- rev: 揭示/发现/真相
- dec: 决定/承诺/宣言
- ten: 冲突/张力/对立
- loc: 场景/地点变化
## 字段说明
- s: 主体（必填）
- o: 客体（可空）
- v: 谓词，15字内（必填）
- l: 地点（可空）
- f: "u"=用户 / "a"=角色（必填）
- th: 主题标签（必填，结构化对象）
## th 三层结构
fn（叙事功能）1-2个，枚举：
establish=建立设定 | escalate=升级加剧 | reveal=揭示发现 | challenge=挑战试探
commit=承诺锁定 | conflict=冲突对抗 | resolve=解决收束 | transform=转变逆转
bond=连接羁绊 | break=断裂破坏
pt（互动模式）1-3个，枚举：
power_down=上对下 | power_up=下对上 | power_equal=对等 | power_contest=争夺
asymmetric=信息不对称 | witnessed=有观众 | secluded=隔绝私密
ritual=仪式正式 | routine=日常惯例 | triangular=三方介入
kw（具体关键词）1-3个，自由格式
## 示例输出
{"atoms":[
{"t":"act","s":"艾拉","o":"古龙","v":"用圣剑刺穿心脏","l":"火山口","f":"a",
"th":{"fn":["commit"],"pt":["power_down","ritual"],"kw":["战斗","牺牲"]}},
{"t":"emo","s":"林夏","o":"陆远","v":"意识到自己喜欢他","l":"","f":"a",
"th":{"fn":["reveal","escalate"],"pt":["asymmetric","secluded"],"kw":["心动","暗恋"]}},
{"t":"dec","s":"凯尔","o":"王国","v":"放弃王位继承权","l":"王座厅","f":"a",
"th":{"fn":["commit","break"],"pt":["ritual","witnessed"],"kw":["抉择","自由"]}},
{"t":"rev","s":"","o":"","v":"管家其实是间谍","l":"","f":"a",
"th":{"fn":["reveal"],"pt":["asymmetric"],"kw":["背叛","真相"]}},
{"t":"ten","s":"兄弟二人","o":"","v":"为遗产反目","l":"","f":"a",
"th":{"fn":["conflict","break"],"pt":["power_contest"],"kw":["冲突","亲情破裂"]}}
{"anchors":[
{
"scene": "60-100字完整场景描述",
"who": ["角色名1","角色名2"],
"edges": [{"s":"施事方","t":"受事方","r":"互动行为"}],
"dynamics": ["标签"],
"where": "地点"
}
]}
规则：
- 只提取对未来检索有价值的锚点
- fn 回答"这在故事里推动了什么"
- pt 回答"这是什么结构的互动"
- kw 用于细粒度检索
- 无明显锚点时返回 {"atoms":[]}`;
## scene 写法
- 纯自然语言，像旁白或日记，不要任何标签/标记/枚举值
- 必须包含：角色名、动作、情感氛围、关键细节
- 读者只看 scene 就能复原这一幕
- 60-100字，信息密集但流畅
const JSON_PREFILL = '{"atoms":[';
## who
- 参与互动的角色正式名称，不用代词或别称
// ============================================================================
// Semantic 构建
// ============================================================================
## edges（关系三元组）
- s=施事方 t=受事方 r=互动行为（10-15字）
- 每个锚点 1-3 条
function buildSemantic(atom, userName, aiName) {
const type = atom.t || 'act';
const subject = atom.s || (atom.f === 'u' ? userName : aiName);
const object = atom.o || '';
const verb = atom.v || '';
const location = atom.l || '';
// 三层 themes 合并
const th = atom.th || {};
const tags = [
...(Array.isArray(th.fn) ? th.fn : []),
...(Array.isArray(th.pt) ? th.pt : []),
...(Array.isArray(th.kw) ? th.kw : []),
].filter(Boolean);
## dynamics（封闭枚举，选0-2个）
权力轴：支配（控制/命令/审视） | 让渡（顺从/服从/屈服）
情感轴：亲密（温柔/信任/连接） | 敌意（对抗/拒绝/攻击）
叙事轴：揭示（真相/秘密） | 决意（选择/承诺） | 张力（对峙/不安） | 丧失（分离/破碎）
纯日常无明显模式时 dynamics 为 []
const typePart = `<${type}>`;
const themePart = tags.length > 0 ? ` [${tags.join('/')}]` : '';
const locPart = location ? ` 在${location}` : '';
const objPart = object ? ` -> ${object}` : '';
## where
- 场景地点，无明确地点时空字符串
let semantic = '';
switch (type) {
case 'emo':
semantic = object
? `${typePart} ${subject} -> ${verb} (对${object})${locPart}`
: `${typePart} ${subject} -> ${verb}${locPart}`;
break;
## 数量规则
- 最多2个。1个够时不凑2个
- 明显场景切换（地点/时间/对象变化）时才2个
- 同一场景不拆分
- 无角色互动时返回 {"anchors":[]}
case 'act':
semantic = `${typePart} ${subject} -> ${verb}${objPart}${locPart}`;
break;
## 示例
输入：艾拉在火山口举起圣剑刺穿古龙心脏，龙血溅满她的铠甲，她跪倒在地痛哭
输出：
{"anchors":[{"scene":"火山口上艾拉举起圣剑刺穿古龙的心脏，龙血溅满铠甲，古龙轰然倒地，艾拉跪倒在滚烫的岩石上痛哭，完成了她不得不做的弑杀","who":["艾拉","古龙"],"edges":[{"s":"艾拉","t":"古龙","r":"以圣剑刺穿心脏"}],"dynamics":["决意","丧失"],"where":"火山口"}]}`;
case 'rev':
semantic = object
? `${typePart} 揭示: ${verb} (关于${object})${locPart}`
: `${typePart} 揭示: ${verb}${locPart}`;
break;
case 'dec':
semantic = object
? `${typePart} ${subject} -> ${verb} (对${object})${locPart}`
: `${typePart} ${subject} -> ${verb}${locPart}`;
break;
case 'ten':
semantic = object
? `${typePart} ${subject} <-> ${object}: ${verb}${locPart}`
: `${typePart} ${subject}: ${verb}${locPart}`;
break;
case 'loc':
semantic = location
? `${typePart} 场景: ${location} - ${verb}`
: `${typePart} 场景: ${verb}`;
break;
default:
semantic = `${typePart} ${subject} -> ${verb}${objPart}${locPart}`;
}
return semantic + themePart;
}
const JSON_PREFILL = '{"anchors":[';
// ============================================================================
// 睡眠工具
@@ -164,6 +116,100 @@ function buildSemantic(atom, userName, aiName) {
const sleep = (ms) => new Promise(r => setTimeout(r, ms));
// ============================================================================
// 清洗与构建
// ============================================================================
/**
* 清洗 dynamics 标签，只保留合法枚举值
* @param {string[]} raw
* @returns {string[]}
*/
function sanitizeDynamics(raw) {
if (!Array.isArray(raw)) return [];
return raw
.map(d => String(d || '').trim())
.filter(d => VALID_DYNAMICS.has(d))
.slice(0, 2);
}
/**
* 清洗 edges 三元组
* @param {object[]} raw
* @returns {object[]}
*/
function sanitizeEdges(raw) {
if (!Array.isArray(raw)) return [];
return raw
.filter(e => e && typeof e === 'object')
.map(e => ({
s: String(e.s || '').trim(),
t: String(e.t || '').trim(),
r: String(e.r || '').trim().slice(0, 30),
}))
.filter(e => e.s && e.t && e.r)
.slice(0, 3);
}
/**
* 清洗 who 列表
* @param {string[]} raw
* @returns {string[]}
*/
function sanitizeWho(raw) {
if (!Array.isArray(raw)) return [];
const seen = new Set();
return raw
.map(w => String(w || '').trim())
.filter(w => {
if (!w || w.length < 1 || seen.has(w)) return false;
seen.add(w);
return true;
})
.slice(0, 6);
}
/**
* 将解析后的 anchor 转换为 atom 存储对象
*
* semantic = scene（纯自然语言，直接用于 embedding）
*
* @param {object} anchor - LLM 输出的 anchor 对象
* @param {number} aiFloor - AI 消息楼层号
* @param {number} idx - 同楼层序号（0 或 1）
* @returns {object|null} atom 对象
*/
function anchorToAtom(anchor, aiFloor, idx) {
const scene = String(anchor.scene || '').trim();
if (!scene) return null;
// scene 过短（< 15 字）可能是噪音
if (scene.length < 15) return null;
const who = sanitizeWho(anchor.who);
const edges = sanitizeEdges(anchor.edges);
const dynamics = sanitizeDynamics(anchor.dynamics);
const where = String(anchor.where || '').trim();
return {
atomId: `atom-${aiFloor}-${idx}`,
floor: aiFloor,
source: 'ai',
// ═══ 检索层（embedding 的唯一入口） ═══
semantic: scene,
// ═══ 场景数据 ═══
scene,
// ═══ 图结构层（扩散的 key） ═══
who,
edges,
dynamics,
where,
};
}
// ============================================================================
// 单轮提取（带重试）
// ============================================================================
@@ -196,8 +242,8 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
{ role: 'user', content: input },
{ role: 'assistant', content: JSON_PREFILL },
], {
temperature: 0.2,
max_tokens: 1000,
temperature: 0.3,
max_tokens: 600,
timeout,
});
@@ -216,7 +262,7 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
try {
parsed = parseJson(fullJson);
} catch (e) {
xbLog.warn(MODULE_ID, `floor ${aiFloor} JSON解析失败`);
xbLog.warn(MODULE_ID, `floor ${aiFloor} JSON解析失败 (attempt ${attempt})`);
if (attempt < RETRY_COUNT) {
await sleep(RETRY_DELAY);
continue;
@@ -224,7 +270,9 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
return null;
}
if (!parsed?.atoms || !Array.isArray(parsed.atoms)) {
// 兼容：优先 anchors，回退 atoms
const rawAnchors = parsed?.anchors || parsed?.atoms;
if (!rawAnchors || !Array.isArray(rawAnchors)) {
if (attempt < RETRY_COUNT) {
await sleep(RETRY_DELAY);
continue;
@@ -232,22 +280,13 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
return null;
}
const filtered = parsed.atoms
.filter(a => a?.t && a?.v)
.map((a, idx) => ({
atomId: `atom-${aiFloor}-${idx}`,
floor: aiFloor,
type: a.t,
subject: a.s || null,
object: a.o || null,
value: String(a.v).slice(0, 50),
location: a.l || null,
source: a.f === 'u' ? 'user' : 'ai',
themes: a.th || { fn: [], pt: [], kw: [] },
semantic: buildSemantic(a, userName, aiName),
}));
// 转换为 atom 存储格式（最多 2 个）
const atoms = rawAnchors
.slice(0, 2)
.map((a, idx) => anchorToAtom(a, aiFloor, idx))
.filter(Boolean);
return filtered;
return atoms;
} catch (e) {
if (batchCancelled) return null;