mirror of
https://github.com/Youzini-afk/ST-Bionic-Memory-Ecology.git
synced 2026-05-15 22:30:38 +08:00
267 lines
9.2 KiB
JavaScript
267 lines
9.2 KiB
JavaScript
// ST-BME: 层级压缩引擎
|
||
// 超过阈值的节点被 LLM 总结为更高层级的压缩节点
|
||
|
||
import { createNode, addNode, createEdge, addEdge, getActiveNodes, getNode } from './graph.js';
|
||
import { callLLMForJSON } from './llm.js';
|
||
import { embedText } from './embedding.js';
|
||
import { isDirectVectorConfig } from './vector-index.js';
|
||
|
||
function createAbortError(message = '操作已终止') {
|
||
const error = new Error(message);
|
||
error.name = 'AbortError';
|
||
return error;
|
||
}
|
||
|
||
function throwIfAborted(signal) {
|
||
if (signal?.aborted) {
|
||
throw signal.reason instanceof Error ? signal.reason : createAbortError();
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 对指定类型执行层级压缩
|
||
*
|
||
* @param {object} params
|
||
* @param {object} params.graph - 当前图状态
|
||
* @param {object} params.typeDef - 要压缩的类型定义
|
||
* @param {object} params.embeddingConfig - Embedding API 配置
|
||
* @param {boolean} [params.force=false] - 忽略阈值强制压缩
|
||
* @returns {Promise<{created: number, archived: number}>}
|
||
*/
|
||
export async function compressType({ graph, typeDef, embeddingConfig, force = false, customPrompt, signal }) {
|
||
const compression = typeDef.compression;
|
||
if (!compression || compression.mode !== 'hierarchical') {
|
||
return { created: 0, archived: 0 };
|
||
}
|
||
|
||
let totalCreated = 0;
|
||
let totalArchived = 0;
|
||
|
||
// 从最低层级开始逐层压缩
|
||
for (let level = 0; level < compression.maxDepth; level++) {
|
||
throwIfAborted(signal);
|
||
const result = await compressLevel({
|
||
graph,
|
||
typeDef,
|
||
level,
|
||
embeddingConfig,
|
||
force,
|
||
customPrompt,
|
||
signal,
|
||
});
|
||
|
||
totalCreated += result.created;
|
||
totalArchived += result.archived;
|
||
|
||
// 如果这一层没有压缩发生，停止
|
||
if (result.created === 0) break;
|
||
}
|
||
|
||
return { created: totalCreated, archived: totalArchived };
|
||
}
|
||
|
||
/**
|
||
* 压缩特定层级的节点
|
||
*/
|
||
async function compressLevel({ graph, typeDef, level, embeddingConfig, force, customPrompt, signal }) {
|
||
const compression = typeDef.compression;
|
||
throwIfAborted(signal);
|
||
|
||
// 获取该层级的活跃叶子节点
|
||
const levelNodes = getActiveNodes(graph, typeDef.id)
|
||
.filter(n => n.level === level)
|
||
.sort((a, b) => a.seq - b.seq);
|
||
|
||
const threshold = force ? Math.max(2, compression.fanIn) : compression.threshold;
|
||
const keepRecent = force ? 0 : compression.keepRecentLeaves;
|
||
|
||
// 不够阈值，无需压缩
|
||
if (levelNodes.length <= threshold) {
|
||
return { created: 0, archived: 0 };
|
||
}
|
||
|
||
// 排除最近的节点
|
||
const compressible = levelNodes.slice(0, levelNodes.length - keepRecent);
|
||
if (compressible.length < compression.fanIn) {
|
||
return { created: 0, archived: 0 };
|
||
}
|
||
|
||
let created = 0;
|
||
let archived = 0;
|
||
|
||
// 按 fanIn 分组压缩
|
||
for (let i = 0; i < compressible.length; i += compression.fanIn) {
|
||
const batch = compressible.slice(i, i + compression.fanIn);
|
||
if (batch.length < 2) break; // 至少 2 个才压缩
|
||
|
||
// 调用 LLM 总结
|
||
const summaryResult = await summarizeBatch(batch, typeDef, customPrompt, signal);
|
||
if (!summaryResult) continue;
|
||
|
||
// 创建压缩节点
|
||
const compressedNode = createNode({
|
||
type: typeDef.id,
|
||
fields: summaryResult.fields,
|
||
seq: batch[batch.length - 1].seq,
|
||
seqRange: [batch[0].seqRange?.[0] ?? batch[0].seq, batch[batch.length - 1].seqRange?.[1] ?? batch[batch.length - 1].seq],
|
||
importance: Math.max(...batch.map(n => n.importance)),
|
||
});
|
||
|
||
compressedNode.level = level + 1;
|
||
compressedNode.childIds = batch.map(n => n.id);
|
||
|
||
// 生成 embedding
|
||
if (isDirectVectorConfig(embeddingConfig) && summaryResult.fields.summary) {
|
||
const vec = await embedText(summaryResult.fields.summary, embeddingConfig, { signal });
|
||
if (vec) compressedNode.embedding = Array.from(vec);
|
||
}
|
||
|
||
addNode(graph, compressedNode);
|
||
migrateBatchEdges(graph, batch, compressedNode);
|
||
created++;
|
||
|
||
// 归档子节点
|
||
for (const child of batch) {
|
||
child.archived = true;
|
||
child.parentId = compressedNode.id;
|
||
archived++;
|
||
}
|
||
}
|
||
|
||
return { created, archived };
|
||
}
|
||
|
||
function migrateBatchEdges(graph, batch, compressedNode) {
|
||
const batchIds = new Set(batch.map(node => node.id));
|
||
const activeNodeIds = new Set(getActiveNodes(graph).map(node => node.id));
|
||
|
||
for (const edge of graph.edges) {
|
||
if (edge.invalidAt || edge.expiredAt) continue;
|
||
|
||
const fromInside = batchIds.has(edge.fromId);
|
||
const toInside = batchIds.has(edge.toId);
|
||
if (!fromInside && !toInside) continue;
|
||
if (fromInside && toInside) continue;
|
||
|
||
const newFromId = fromInside ? compressedNode.id : edge.fromId;
|
||
const newToId = toInside ? compressedNode.id : edge.toId;
|
||
|
||
if (newFromId === newToId) continue;
|
||
if (!activeNodeIds.has(newFromId) || !activeNodeIds.has(newToId)) continue;
|
||
if (!getNode(graph, newFromId) || !getNode(graph, newToId)) continue;
|
||
|
||
const migratedEdge = createEdge({
|
||
fromId: newFromId,
|
||
toId: newToId,
|
||
relation: edge.relation,
|
||
strength: edge.strength,
|
||
edgeType: edge.edgeType,
|
||
});
|
||
migratedEdge.validAt = edge.validAt ?? migratedEdge.validAt;
|
||
migratedEdge.invalidAt = edge.invalidAt ?? migratedEdge.invalidAt;
|
||
migratedEdge.expiredAt = edge.expiredAt ?? migratedEdge.expiredAt;
|
||
|
||
addEdge(graph, migratedEdge);
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 调用 LLM 总结一批节点
|
||
*/
|
||
async function summarizeBatch(nodes, typeDef, customPrompt, signal) {
|
||
const nodeDescriptions = nodes.map((n, i) => {
|
||
const fieldsStr = Object.entries(n.fields)
|
||
.filter(([_, v]) => v)
|
||
.map(([k, v]) => `${k}: ${v}`)
|
||
.join('\n ');
|
||
return `节点 ${i + 1} [楼层 ${n.seq}]:\n ${fieldsStr}`;
|
||
}).join('\n\n');
|
||
|
||
const instruction = typeDef.compression.instruction || '将以下节点压缩总结为一条精炼记录。';
|
||
|
||
const systemPrompt = customPrompt || [
|
||
'你是一个记忆压缩器。将多个同类型节点总结为一条更高层级的压缩节点。',
|
||
instruction,
|
||
'',
|
||
'输出格式为严格 JSON：',
|
||
`{"fields": {${typeDef.columns.map(c => `"${c.name}": "..."`).join(', ')}}}`,
|
||
'',
|
||
'规则：',
|
||
'- 保留关键信息：因果关系、不可逆结果、未解决伏笔',
|
||
'- 去除重复和低信息密度内容',
|
||
'- 压缩后文本应精炼，目标 150 字左右',
|
||
].join('\n');
|
||
|
||
const userPrompt = `请压缩以下 ${nodes.length} 个 "${typeDef.label}" 节点：\n\n${nodeDescriptions}`;
|
||
|
||
return await callLLMForJSON({ systemPrompt, userPrompt, maxRetries: 1, signal });
|
||
}
|
||
|
||
/**
|
||
* 对所有支持压缩的类型执行压缩
|
||
*
|
||
* @param {object} graph
|
||
* @param {object[]} schema
|
||
* @param {object} embeddingConfig
|
||
* @param {boolean} [force=false]
|
||
* @returns {Promise<{created: number, archived: number}>}
|
||
*/
|
||
export async function compressAll(graph, schema, embeddingConfig, force = false, customPrompt, signal) {
|
||
let totalCreated = 0;
|
||
let totalArchived = 0;
|
||
|
||
for (const typeDef of schema) {
|
||
throwIfAborted(signal);
|
||
if (typeDef.compression?.mode === 'hierarchical') {
|
||
const result = await compressType({ graph, typeDef, embeddingConfig, force, customPrompt, signal });
|
||
totalCreated += result.created;
|
||
totalArchived += result.archived;
|
||
}
|
||
}
|
||
|
||
return { created: totalCreated, archived: totalArchived };
|
||
}
|
||
|
||
// ==================== v2: 主动遗忘（SleepGate 启发） ====================
|
||
|
||
/**
|
||
* 睡眠清理周期
|
||
* 评估每个节点的保留价值，低于阈值的归档（遗忘）
|
||
*
|
||
* @param {object} graph - 图状态
|
||
* @param {object} settings - 包含 forgetThreshold 的设置
|
||
* @returns {{forgotten: number}} 本次遗忘的节点数
|
||
*/
|
||
export function sleepCycle(graph, settings) {
|
||
const threshold = settings.forgetThreshold ?? 0.5;
|
||
const nodes = getActiveNodes(graph);
|
||
const now = Date.now();
|
||
let forgotten = 0;
|
||
|
||
for (const node of nodes) {
|
||
// 跳过常驻类型（synopsis, rule 等重要节点不应被遗忘）
|
||
if (node.type === 'synopsis' || node.type === 'rule' || node.type === 'thread') continue;
|
||
// 跳过高重要性节点
|
||
if (node.importance >= 8) continue;
|
||
// 跳过最近创建的节点（< 1 小时）
|
||
if (now - node.createdTime < 3600000) continue;
|
||
|
||
// 计算保留价值 = importance × recency × (1 + accessFreq)
|
||
const ageHours = (now - node.createdTime) / 3600000;
|
||
const recency = 1 / (1 + Math.log10(1 + ageHours));
|
||
const accessFreq = node.accessCount / Math.max(1, ageHours / 24);
|
||
const retentionValue = (node.importance / 10) * recency * (1 + accessFreq);
|
||
|
||
if (retentionValue < threshold) {
|
||
node.archived = true;
|
||
forgotten++;
|
||
}
|
||
}
|
||
|
||
if (forgotten > 0) {
|
||
console.log(`[ST-BME] 主动遗忘: ${forgotten} 个低价值节点已归档`);
|
||
}
|
||
|
||
return { forgotten };
|
||
}
|