mirror of
https://github.com/Youzini-afk/ST-Bionic-Memory-Ecology.git
synced 2026-05-15 22:30:38 +08:00
525 lines
14 KiB
JavaScript
525 lines
14 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 { embedText } from "./embedding.js";
|
||
import {
|
||
addEdge,
|
||
addNode,
|
||
createEdge,
|
||
createNode,
|
||
getActiveNodes,
|
||
getNode,
|
||
} from "./graph.js";
|
||
import { callLLMForJSON } from "./llm.js";
|
||
import {
|
||
getScopeOwnerKey,
|
||
getScopeRegionKey,
|
||
normalizeMemoryScope,
|
||
} from "./memory-scope.js";
|
||
import {
|
||
buildTaskExecutionDebugContext,
|
||
buildTaskLlmPayload,
|
||
buildTaskPrompt,
|
||
} from "./prompt-builder.js";
|
||
import { getSTContextForPrompt } from "./st-context.js";
|
||
import { applyTaskRegex } from "./task-regex.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 createTaskLlmDebugContext(promptBuild, regexInput) {
|
||
return typeof buildTaskExecutionDebugContext === "function"
|
||
? buildTaskExecutionDebugContext(promptBuild, { regexInput })
|
||
: null;
|
||
}
|
||
|
||
function resolveTaskPromptPayload(promptBuild, fallbackUserPrompt = "") {
|
||
if (typeof buildTaskLlmPayload === "function") {
|
||
return buildTaskLlmPayload(promptBuild, fallbackUserPrompt);
|
||
}
|
||
|
||
return {
|
||
systemPrompt: String(promptBuild?.systemPrompt || ""),
|
||
userPrompt: String(fallbackUserPrompt || ""),
|
||
promptMessages: [],
|
||
additionalMessages: Array.isArray(promptBuild?.privateTaskMessages)
|
||
? promptBuild.privateTaskMessages
|
||
: [],
|
||
};
|
||
}
|
||
|
||
function throwIfAborted(signal) {
|
||
if (signal?.aborted) {
|
||
throw signal.reason instanceof Error ? signal.reason : createAbortError();
|
||
}
|
||
}
|
||
|
||
function resolveCompressionWindow(compression = {}, force = false) {
|
||
const fanIn = Number.isFinite(Number(compression?.fanIn))
|
||
? Math.max(2, Number(compression.fanIn))
|
||
: 2;
|
||
const threshold = force
|
||
? fanIn
|
||
: Number.isFinite(Number(compression?.threshold))
|
||
? Math.max(2, Number(compression.threshold))
|
||
: fanIn;
|
||
const keepRecent = force
|
||
? 0
|
||
: Number.isFinite(Number(compression?.keepRecentLeaves))
|
||
? Math.max(0, Number(compression.keepRecentLeaves))
|
||
: 0;
|
||
|
||
return {
|
||
fanIn,
|
||
threshold,
|
||
keepRecent,
|
||
};
|
||
}
|
||
|
||
/**
|
||
* 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,
|
||
settings = {},
|
||
}) {
|
||
const compression = typeDef.compression;
|
||
if (!compression || compression.mode !== "hierarchical") {
|
||
return { created: 0, archived: 0 };
|
||
}
|
||
const maxDepth = Number.isFinite(Number(compression.maxDepth))
|
||
? Math.max(1, Number(compression.maxDepth))
|
||
: 1;
|
||
|
||
let totalCreated = 0;
|
||
let totalArchived = 0;
|
||
|
||
// 从最低层级开始逐层压缩
|
||
for (let level = 0; level < maxDepth; level++) {
|
||
throwIfAborted(signal);
|
||
const result = await compressLevel({
|
||
graph,
|
||
typeDef,
|
||
level,
|
||
embeddingConfig,
|
||
force,
|
||
customPrompt,
|
||
signal,
|
||
settings,
|
||
});
|
||
|
||
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,
|
||
settings = {},
|
||
}) {
|
||
const compression = typeDef.compression;
|
||
const { fanIn, threshold, keepRecent } = resolveCompressionWindow(
|
||
compression,
|
||
force,
|
||
);
|
||
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);
|
||
let created = 0;
|
||
let archived = 0;
|
||
|
||
for (const group of groupCompressionCandidates(levelNodes)) {
|
||
if (force ? group.length < fanIn : group.length <= threshold) {
|
||
continue;
|
||
}
|
||
|
||
const compressible = group.slice(0, Math.max(0, group.length - keepRecent));
|
||
if (compressible.length < fanIn) {
|
||
continue;
|
||
}
|
||
|
||
for (let i = 0; i < compressible.length; i += fanIn) {
|
||
const batch = compressible.slice(i, i + fanIn);
|
||
if (batch.length < 2) break;
|
||
|
||
const summaryResult = await summarizeBatch(
|
||
batch,
|
||
typeDef,
|
||
customPrompt,
|
||
signal,
|
||
settings,
|
||
);
|
||
if (!summaryResult) continue;
|
||
|
||
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)),
|
||
scope: normalizeMemoryScope(batch[0]?.scope),
|
||
});
|
||
|
||
compressedNode.level = level + 1;
|
||
compressedNode.childIds = batch.map((n) => n.id);
|
||
|
||
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++;
|
||
|
||
for (const child of batch) {
|
||
child.archived = true;
|
||
child.parentId = compressedNode.id;
|
||
archived++;
|
||
}
|
||
}
|
||
}
|
||
|
||
return { created, archived };
|
||
}
|
||
|
||
function groupCompressionCandidates(nodes = []) {
|
||
const groups = new Map();
|
||
for (const node of nodes) {
|
||
const normalizedScope = normalizeMemoryScope(node?.scope);
|
||
const key =
|
||
normalizedScope.layer === "pov"
|
||
? [
|
||
"pov",
|
||
getScopeOwnerKey(normalizedScope) || "owner:none",
|
||
node.type || "",
|
||
].join("::")
|
||
: [
|
||
"objective",
|
||
getScopeRegionKey(normalizedScope) || "region:global",
|
||
node.type || "",
|
||
].join("::");
|
||
if (!groups.has(key)) {
|
||
groups.set(key, []);
|
||
}
|
||
groups.get(key).push(node);
|
||
}
|
||
return [...groups.values()].map((group) =>
|
||
group.sort((a, b) => a.seq - b.seq),
|
||
);
|
||
}
|
||
|
||
function inspectCompressibleGroup(group = [], compression = {}, force = false) {
|
||
const { fanIn, threshold, keepRecent } = resolveCompressionWindow(
|
||
compression,
|
||
force,
|
||
);
|
||
if (force ? group.length < fanIn : group.length <= threshold) {
|
||
return null;
|
||
}
|
||
|
||
const compressible = group.slice(0, Math.max(0, group.length - keepRecent));
|
||
if (compressible.length < fanIn) {
|
||
return null;
|
||
}
|
||
|
||
return {
|
||
candidateCount: compressible.length,
|
||
fanIn,
|
||
threshold,
|
||
keepRecent,
|
||
};
|
||
}
|
||
|
||
export function inspectAutoCompressionCandidates(
|
||
graph,
|
||
schema = [],
|
||
force = false,
|
||
) {
|
||
const safeSchema = Array.isArray(schema) ? schema : [];
|
||
for (const typeDef of safeSchema) {
|
||
if (typeDef?.compression?.mode !== "hierarchical") continue;
|
||
const maxDepth = Number.isFinite(Number(typeDef?.compression?.maxDepth))
|
||
? Math.max(1, Number(typeDef.compression.maxDepth))
|
||
: 1;
|
||
|
||
for (let level = 0; level < maxDepth; level++) {
|
||
const levelNodes = getActiveNodes(graph, typeDef.id)
|
||
.filter((node) => Number(node?.level || 0) === level)
|
||
.sort((a, b) => a.seq - b.seq);
|
||
|
||
for (const group of groupCompressionCandidates(levelNodes)) {
|
||
const summary = inspectCompressibleGroup(
|
||
group,
|
||
typeDef.compression,
|
||
force,
|
||
);
|
||
if (!summary) continue;
|
||
return {
|
||
hasCandidates: true,
|
||
typeId: String(typeDef.id || ""),
|
||
level,
|
||
candidateCount: summary.candidateCount,
|
||
threshold: summary.threshold,
|
||
fanIn: summary.fanIn,
|
||
keepRecent: summary.keepRecent,
|
||
reason: "",
|
||
};
|
||
}
|
||
}
|
||
}
|
||
|
||
return {
|
||
hasCandidates: false,
|
||
typeId: "",
|
||
level: null,
|
||
candidateCount: 0,
|
||
threshold: 0,
|
||
fanIn: 0,
|
||
keepRecent: 0,
|
||
reason: "已到自动压缩周期，但当前没有达到内部压缩阈值的候选组",
|
||
};
|
||
}
|
||
|
||
function migrateBatchEdges(graph, batch, compressedNode) {
|
||
const batchIds = new Set(batch.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 (!getNode(graph, newFromId) || !getNode(graph, newToId)) continue;
|
||
|
||
const migratedEdge = createEdge({
|
||
fromId: newFromId,
|
||
toId: newToId,
|
||
relation: edge.relation,
|
||
strength: edge.strength,
|
||
edgeType: edge.edgeType,
|
||
scope: edge.scope,
|
||
});
|
||
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,
|
||
settings = {},
|
||
) {
|
||
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 ${fieldsStr}`;
|
||
})
|
||
.join("\n\n");
|
||
|
||
const instruction =
|
||
typeDef.compression.instruction || "将以下节点压缩总结为一条精炼记录。";
|
||
|
||
const compressPromptBuild = await buildTaskPrompt(settings, "compress", {
|
||
taskName: "compress",
|
||
nodeContent: nodeDescriptions,
|
||
candidateNodes: nodeDescriptions,
|
||
currentRange: `${nodes[0]?.seq ?? "?"} ~ ${nodes[nodes.length - 1]?.seq ?? "?"}`,
|
||
graphStats: `node_count=${nodes.length}, node_type=${typeDef.id}`,
|
||
...getSTContextForPrompt(),
|
||
});
|
||
const compressRegexInput = { entries: [] };
|
||
const systemPrompt = applyTaskRegex(
|
||
settings,
|
||
"compress",
|
||
"finalPrompt",
|
||
compressPromptBuild.systemPrompt ||
|
||
customPrompt ||
|
||
[
|
||
"你是一个记忆压缩器。将多个同类型节点总结为一条更高层级的压缩节点。",
|
||
instruction,
|
||
"",
|
||
"输出格式为严格 JSON：",
|
||
`{"fields": {${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"),
|
||
compressRegexInput,
|
||
"system",
|
||
);
|
||
|
||
const userPrompt = `请压缩以下 ${nodes.length} 个 "${typeDef.label}" 节点：\n\n${nodeDescriptions}`;
|
||
const promptPayload = resolveTaskPromptPayload(
|
||
compressPromptBuild,
|
||
userPrompt,
|
||
);
|
||
const llmSystemPrompt =
|
||
Array.isArray(promptPayload.promptMessages) &&
|
||
promptPayload.promptMessages.length > 0
|
||
? String(promptPayload.systemPrompt || "")
|
||
: String(promptPayload.systemPrompt || systemPrompt || "");
|
||
|
||
return await callLLMForJSON({
|
||
systemPrompt: llmSystemPrompt,
|
||
userPrompt: promptPayload.userPrompt,
|
||
maxRetries: 1,
|
||
signal,
|
||
taskType: "compress",
|
||
debugContext: createTaskLlmDebugContext(
|
||
compressPromptBuild,
|
||
compressRegexInput,
|
||
),
|
||
promptMessages: promptPayload.promptMessages,
|
||
additionalMessages: promptPayload.additionalMessages,
|
||
});
|
||
}
|
||
|
||
/**
|
||
* 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,
|
||
settings = {},
|
||
) {
|
||
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,
|
||
settings,
|
||
});
|
||
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 };
|
||
}
|