feat: LLM intent router, direct chat execution, and Memory V2 light intervention
Wire chat intent routing with direct_chat on regular sessions, skill-selected short-circuit to Agent, memory light/heavy intervention tiers, and fix direct chat UI stuck streaming after completion. Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
john
@@ -0,0 +1,772 @@
|
||||
import {
|
||||
buildChatSkillPrompt,
|
||||
extractSelectedChatSkillName,
|
||||
hasExplicitChatSkillPrompt,
|
||||
} from './chat-skills.mjs';
|
||||
import { isDirectChatSessionId } from './direct-chat-service.mjs';
|
||||
import {
|
||||
memoryLimitForIntervention,
|
||||
resolveMemoryInterventionMode,
|
||||
} from './memory-intervention.mjs';
|
||||
|
||||
export const CHAT_INTENT_ROUTE = {
|
||||
DIRECT_CHAT: 'direct_chat',
|
||||
AGENT: 'agent_orchestration',
|
||||
};
|
||||
|
||||
const AGENT_ORCHESTRATION_HEADER = '【Memind 任务编排】';
|
||||
|
||||
const SKILL_PROMPT_KEYS = {
|
||||
web: 'web',
|
||||
search: 'search',
|
||||
'static-page-publish': 'generate-page',
|
||||
'form-builder': 'form-builder',
|
||||
'table-viewer': 'table-viewer',
|
||||
'product-campaign-page': 'product-campaign-page',
|
||||
};
|
||||
|
||||
const OBVIOUS_AGENT_PATTERNS = [
|
||||
/\bpublic\/[^\s"'<>]+\.html\b/i,
|
||||
/MindSpace\/[^/\s]+\/public\/[^\s"'<>]+\.html/i,
|
||||
/(?:生成|创建|制作|做|写|设计|发布).{0,24}(?:H5|h5|HTML|html|网页|页面|活动页|宣传页|落地页|分享页)/u,
|
||||
/(?:H5|h5|HTML|html|网页|页面).{0,24}(?:生成|创建|制作|做|写|设计|发布)/u,
|
||||
];
|
||||
|
||||
const OBVIOUS_DIRECT_PATTERNS = [
|
||||
/^(?:你好|您好|在吗|在不在|嗨|hi|hello|hey)[!！。.\s]*$/iu,
|
||||
/^(?:测试\s*\d*|test\s*\d*)[!！。.\s]*$/iu,
|
||||
/^[?？]+$/u,
|
||||
];
|
||||
|
||||
const MEMORY_RECALL_PATTERNS = [
|
||||
/你(?:还)?记得(?:我)?/u,
|
||||
/(?:有没有|是否).{0,6}记住/u,
|
||||
/你对(?:我)?(?:的)?记忆/u,
|
||||
/(?:我|之前).{0,16}(?:说过|提到|聊过|告诉)/u,
|
||||
/(?:我的|之前的)(?:记忆|偏好|计划|目标|想法)/u,
|
||||
/之前(?:说|提|聊)(?:过|的)/u,
|
||||
];
|
||||
|
||||
export function isMemoryRecallQuestion(text) {
|
||||
const normalized = String(text ?? '').trim();
|
||||
if (!normalized) return false;
|
||||
return MEMORY_RECALL_PATTERNS.some((pattern) => pattern.test(normalized));
|
||||
}
|
||||
|
||||
const DEFAULT_ROUTER_TIMEOUT_MS = 1500;
|
||||
const DEFAULT_ROUTER_MEMORY_LIMIT = 8;
|
||||
const DEFAULT_ROUTER_MIN_CONFIDENCE = 0.55;
|
||||
|
||||
function envFlag(value, fallback = false) {
|
||||
const raw = String(value ?? '').trim().toLowerCase();
|
||||
if (!raw) return fallback;
|
||||
return ['1', 'true', 'yes', 'on'].includes(raw);
|
||||
}
|
||||
|
||||
function boundedNumber(value, fallback, { min = 0, max = Number.POSITIVE_INFINITY } = {}) {
|
||||
if (value == null || value === '') return fallback;
|
||||
const parsed = Number(value);
|
||||
if (!Number.isFinite(parsed)) return fallback;
|
||||
return Math.min(max, Math.max(min, parsed));
|
||||
}
|
||||
|
||||
function pickDefined(source, keys) {
|
||||
return Object.fromEntries(
|
||||
keys
|
||||
.filter((key) => source[key] !== undefined)
|
||||
.map((key) => [key, source[key]]),
|
||||
);
|
||||
}
|
||||
|
||||
function truncateText(value, maxLength) {
|
||||
const text = String(value ?? '').replace(/\s+/g, ' ').trim();
|
||||
if (!text) return '';
|
||||
return text.length > maxLength ? `${text.slice(0, maxLength)}...` : text;
|
||||
}
|
||||
|
||||
function firstSentence(value, maxLength) {
|
||||
const text = truncateText(value, maxLength);
|
||||
if (!text) return '';
|
||||
const match = text.match(/^(.{12,}?[。.!！?？])/u);
|
||||
return truncateText(match?.[1] ?? text, maxLength);
|
||||
}
|
||||
|
||||
function normalizeMemoryText(item) {
|
||||
if (typeof item === 'string') return { label: null, text: item };
|
||||
const text = item?.text ?? item?.memory_text ?? item?.memoryText ?? item?.content ?? item?.summary ?? '';
|
||||
const label = String(item?.label ?? item?.type ?? '').trim() || null;
|
||||
return { label, text };
|
||||
}
|
||||
|
||||
function withTimeout(promise, timeoutMs, label) {
|
||||
const timeout = Number(timeoutMs);
|
||||
if (!Number.isFinite(timeout) || timeout <= 0) return promise;
|
||||
let timer = null;
|
||||
const timeoutPromise = new Promise((_, reject) => {
|
||||
timer = setTimeout(() => {
|
||||
const err = new Error(`${label} timed out after ${timeout}ms`);
|
||||
err.code = 'CHAT_INTENT_ROUTER_TIMEOUT';
|
||||
reject(err);
|
||||
}, timeout);
|
||||
});
|
||||
return Promise.race([promise, timeoutPromise]).finally(() => {
|
||||
if (timer) clearTimeout(timer);
|
||||
});
|
||||
}
|
||||
|
||||
function messageDisplayText(message) {
|
||||
const displayText = message?.metadata?.displayText;
|
||||
if (typeof displayText === 'string' && displayText.trim()) return displayText.trim();
|
||||
return extractMessageText(message);
|
||||
}
|
||||
|
||||
function extractMessageText(message) {
|
||||
const content = message?.content;
|
||||
if (typeof content === 'string') return content.trim();
|
||||
if (!Array.isArray(content)) return String(message?.text ?? message?.value ?? '').trim();
|
||||
return content
|
||||
.map((item) => {
|
||||
if (typeof item === 'string') return item;
|
||||
if (item?.type === 'text') return item.text ?? '';
|
||||
return '';
|
||||
})
|
||||
.join('\n')
|
||||
.trim();
|
||||
}
|
||||
|
||||
function isTextOnlyUserMessage(message) {
|
||||
const content = message?.content;
|
||||
if (!Array.isArray(content)) return Boolean(extractMessageText(message));
|
||||
if (content.length === 0) return false;
|
||||
return content.every((item) => {
|
||||
if (typeof item === 'string') return true;
|
||||
return item?.type === 'text';
|
||||
});
|
||||
}
|
||||
|
||||
function readSelectedChatSkillId(message) {
|
||||
const selected =
|
||||
message?.metadata?.memindRun?.selectedChatSkill ??
|
||||
message?.metadata?.selectedChatSkill;
|
||||
return typeof selected === 'string' && selected.trim() ? selected.trim() : null;
|
||||
}
|
||||
|
||||
function messageHasExplicitChatSkillSelection(message) {
|
||||
if (!message) return false;
|
||||
if (readSelectedChatSkillId(message)) return true;
|
||||
return [extractMessageText(message), messageDisplayText(message)].some((text) =>
|
||||
hasExplicitChatSkillPrompt(text),
|
||||
);
|
||||
}
|
||||
|
||||
function buildSkillSelectionClassification(userMessage) {
|
||||
const selectedSkillId = readSelectedChatSkillId(userMessage);
|
||||
const suggestedSkill =
|
||||
selectedSkillId ??
|
||||
extractSelectedChatSkillName(extractMessageText(userMessage)) ??
|
||||
extractSelectedChatSkillName(messageDisplayText(userMessage));
|
||||
return normalizeClassification({
|
||||
route: CHAT_INTENT_ROUTE.AGENT,
|
||||
confidence: 1,
|
||||
reason: '用户已选择 skill',
|
||||
suggested_skill: suggestedSkill,
|
||||
agent_brief: suggestedSkill ? `使用 ${suggestedSkill} 执行用户任务` : '执行用户选择的 skill 任务',
|
||||
}, { source: 'rule' });
|
||||
}
|
||||
|
||||
function buildRouterSystemPrompt(grantedSkills = []) {
|
||||
const skills = Array.isArray(grantedSkills) ? grantedSkills.filter(Boolean) : [];
|
||||
return [
|
||||
'你是 TKMind H5 聊天意图路由器，只负责判断用户消息应该走哪条处理通道。',
|
||||
'',
|
||||
'系统有两条通道：',
|
||||
'1. direct_chat — 纯文字交流：问答、解释、闲聊、总结已有文字、给建议，不需要写文件、生成页面、调工具。',
|
||||
'2. agent_orchestration — 任务编排 Agent：需要实际执行并产出结果的任务，例如生成/修改 HTML 页面、发布到 MindSpace、搜索实时资料、写代码改仓库、表单收集、数据表格、商品页、docx/长图导出等。',
|
||||
'',
|
||||
'判断原则：',
|
||||
'- 用户只要文字回答，不要求“做出来/发布/生成链接/改文件” → direct_chat',
|
||||
'- 用户要产出可访问页面、文件、链接，或需要工具/skills → agent_orchestration',
|
||||
'- 不确定时优先 agent_orchestration，避免漏执行',
|
||||
'- 记忆线索只用于辅助判断本轮意图，不能替用户扩写新需求',
|
||||
'',
|
||||
skills.length ? `当前用户已授权 skills：${skills.join(', ')}` : '当前用户未授权额外 skills。',
|
||||
skills.length
|
||||
? '若走 agent_orchestration，可在 suggested_skill 中填写最匹配的 skill 名称（须来自上述列表），否则填 null。'
|
||||
: 'suggested_skill 通常填 null。',
|
||||
'',
|
||||
'只输出 JSON，不要 markdown，不要解释：',
|
||||
'{"route":"direct_chat|agent_orchestration","confidence":0.0,"reason":"一句话","suggested_skill":null,"agent_brief":"给 Agent 的执行要点，direct_chat 时可为空"}',
|
||||
].join('\n');
|
||||
}
|
||||
|
||||
export function buildRouterContext(resolveResult, {
|
||||
memoryLimit = 5,
|
||||
semanticLimit = 3,
|
||||
goalLimit = 2,
|
||||
memoryTextLimit = 80,
|
||||
semanticTextLimit = 60,
|
||||
goalTextLimit = 60,
|
||||
behaviorTextLimit = 100,
|
||||
} = {}) {
|
||||
const result = resolveResult && typeof resolveResult === 'object' ? resolveResult : {};
|
||||
const lines = [];
|
||||
let itemsUsed = 0;
|
||||
|
||||
const memories = Array.isArray(result.memories) ? result.memories : [];
|
||||
const memoryLines = memories
|
||||
.map((item) => normalizeMemoryText(item))
|
||||
.map(({ label, text }) => {
|
||||
const clipped = truncateText(text, memoryTextLimit);
|
||||
if (!clipped) return '';
|
||||
return `- ${label ? `[${label}] ` : ''}${clipped}`;
|
||||
})
|
||||
.filter(Boolean)
|
||||
.slice(0, memoryLimit);
|
||||
if (memoryLines.length) {
|
||||
lines.push('相关记忆：', ...memoryLines);
|
||||
itemsUsed += memoryLines.length;
|
||||
}
|
||||
|
||||
const semanticLines = (Array.isArray(result.semanticMemories) ? result.semanticMemories : [])
|
||||
.map((item) => (typeof item === 'string' ? item : normalizeMemoryText(item).text))
|
||||
.map((text) => truncateText(text, semanticTextLimit))
|
||||
.filter(Boolean)
|
||||
.slice(0, semanticLimit)
|
||||
.map((text) => `- ${text}`);
|
||||
if (semanticLines.length) {
|
||||
lines.push('语义线索：', ...semanticLines);
|
||||
itemsUsed += semanticLines.length;
|
||||
}
|
||||
|
||||
const behaviorSummary = firstSentence(result.behaviorSummary, behaviorTextLimit);
|
||||
if (behaviorSummary) {
|
||||
lines.push(`行为摘要：${behaviorSummary}`);
|
||||
itemsUsed += 1;
|
||||
}
|
||||
|
||||
const goals = Array.isArray(result.activeGoals) && result.activeGoals.length
|
||||
? result.activeGoals
|
||||
: (Array.isArray(result.contextGoals) ? result.contextGoals : []);
|
||||
const goalLines = goals
|
||||
.map((text) => truncateText(text, goalTextLimit))
|
||||
.filter(Boolean)
|
||||
.slice(0, goalLimit)
|
||||
.map((text) => `- ${text}`);
|
||||
if (goalLines.length) {
|
||||
lines.push('进行中目标：', ...goalLines);
|
||||
itemsUsed += goalLines.length;
|
||||
}
|
||||
|
||||
return {
|
||||
content: lines.join('\n').trim(),
|
||||
itemsUsed,
|
||||
source: result.source ?? null,
|
||||
degraded: Boolean(result.degraded),
|
||||
skipped: Boolean(result.skipped),
|
||||
reason: result.reason ?? null,
|
||||
};
|
||||
}
|
||||
|
||||
function buildRouterUserPrompt({ text, routerContext }) {
|
||||
const context = String(routerContext ?? '').trim() || '无';
|
||||
return [
|
||||
'[Router Context]',
|
||||
context,
|
||||
'',
|
||||
'[User]',
|
||||
text || '(empty)',
|
||||
].join('\n');
|
||||
}
|
||||
|
||||
function parseRouterJson(reply) {
|
||||
const text = String(reply ?? '').trim();
|
||||
if (!text) return null;
|
||||
const fenced = text.match(/```(?:json)?\s*([\s\S]*?)```/i);
|
||||
const candidate = (fenced?.[1] ?? text).trim();
|
||||
try {
|
||||
return JSON.parse(candidate);
|
||||
} catch {
|
||||
const start = candidate.indexOf('{');
|
||||
const end = candidate.lastIndexOf('}');
|
||||
if (start < 0 || end <= start) return null;
|
||||
try {
|
||||
return JSON.parse(candidate.slice(start, end + 1));
|
||||
} catch {
|
||||
return null;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
function normalizeRoute(value) {
|
||||
const normalized = String(value ?? '').trim().toLowerCase();
|
||||
if (normalized === CHAT_INTENT_ROUTE.DIRECT_CHAT) return CHAT_INTENT_ROUTE.DIRECT_CHAT;
|
||||
if (normalized === CHAT_INTENT_ROUTE.AGENT || normalized === 'agent') return CHAT_INTENT_ROUTE.AGENT;
|
||||
return null;
|
||||
}
|
||||
|
||||
export function resolveChatIntentRouterPolicy({ env = process.env, overrides = {} } = {}) {
|
||||
const fallbackRoute = normalizeRoute(env?.MEMIND_CHAT_ROUTER_FALLBACK_ROUTE) ?? CHAT_INTENT_ROUTE.AGENT;
|
||||
return {
|
||||
enabled: envFlag(env?.MEMIND_CHAT_LLM_ROUTER_ENABLED, false),
|
||||
modelProviderKeyId: String(env?.MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID ?? '').trim() || null,
|
||||
model: String(env?.MEMIND_CHAT_ROUTER_MODEL ?? '').trim() || null,
|
||||
modelApiType: String(env?.MEMIND_CHAT_ROUTER_MODEL_API ?? '').trim() || 'chat',
|
||||
temperature: boundedNumber(env?.MEMIND_CHAT_ROUTER_TEMPERATURE, 0, { min: 0, max: 2 }),
|
||||
minConfidence: boundedNumber(
|
||||
env?.MEMIND_CHAT_ROUTER_MIN_CONFIDENCE,
|
||||
DEFAULT_ROUTER_MIN_CONFIDENCE,
|
||||
{ min: 0, max: 1 },
|
||||
),
|
||||
memoryResolveEnabled: envFlag(env?.MEMIND_CHAT_ROUTER_MEMORY_ENABLED, true),
|
||||
memoryResolveLimit: Math.round(boundedNumber(
|
||||
env?.MEMIND_CHAT_ROUTER_MEMORY_LIMIT,
|
||||
DEFAULT_ROUTER_MEMORY_LIMIT,
|
||||
{ min: 1, max: 50 },
|
||||
)),
|
||||
timeoutMs: Math.round(boundedNumber(
|
||||
env?.MEMIND_CHAT_ROUTER_TIMEOUT_MS,
|
||||
DEFAULT_ROUTER_TIMEOUT_MS,
|
||||
{ min: 0, max: 30_000 },
|
||||
)),
|
||||
fallbackRoute,
|
||||
...overrides,
|
||||
};
|
||||
}
|
||||
|
||||
function normalizeClassification(raw, { source, fallbackRoute = CHAT_INTENT_ROUTE.AGENT } = {}) {
|
||||
const route = normalizeRoute(raw?.route) ?? fallbackRoute;
|
||||
const confidenceRaw = Number(raw?.confidence);
|
||||
const confidence = Number.isFinite(confidenceRaw)
|
||||
? Math.min(1, Math.max(0, confidenceRaw))
|
||||
: source === 'llm' ? 0.7 : 1;
|
||||
const suggestedSkill = String(raw?.suggested_skill ?? raw?.suggestedSkill ?? '').trim() || null;
|
||||
const agentBrief = String(raw?.agent_brief ?? raw?.agentBrief ?? '').trim();
|
||||
const reason = String(raw?.reason ?? '').trim() || (source === 'llm' ? '模型路由判定' : '规则路由判定');
|
||||
return {
|
||||
route,
|
||||
confidence,
|
||||
reason,
|
||||
suggestedSkill,
|
||||
agentBrief,
|
||||
source,
|
||||
providerKeyId: raw?.providerKeyId ?? raw?.provider_key_id ?? null,
|
||||
model: raw?.model ?? null,
|
||||
memory: raw?.memory ?? null,
|
||||
};
|
||||
}
|
||||
|
||||
function resolveSkillPrompt(suggestedSkill, grantedSkills = []) {
|
||||
const skillName = String(suggestedSkill ?? '').trim();
|
||||
if (!skillName) return '';
|
||||
if (grantedSkills.length > 0 && !grantedSkills.includes(skillName)) return '';
|
||||
const promptKey = SKILL_PROMPT_KEYS[skillName];
|
||||
if (!promptKey) return '';
|
||||
return buildChatSkillPrompt(promptKey, skillName);
|
||||
}
|
||||
|
||||
export function buildAgentOrchestrationAgentText({
|
||||
displayText,
|
||||
classification,
|
||||
skillPrompt = '',
|
||||
}) {
|
||||
const taskBody = String(displayText ?? '').trim();
|
||||
const lines = [
|
||||
`${AGENT_ORCHESTRATION_HEADER}以下为用户任务，请使用工具与技能实际执行并产出结果，不要只做文字描述。`,
|
||||
`路由判定：${classification.reason}`,
|
||||
classification.agentBrief ? `执行要点：${classification.agentBrief}` : '',
|
||||
classification.suggestedSkill ? `建议 skill：${classification.suggestedSkill}` : '',
|
||||
skillPrompt,
|
||||
'',
|
||||
'用户任务：',
|
||||
taskBody,
|
||||
].filter((line, index, all) => line !== '' || index === all.length - 2);
|
||||
return lines.join('\n');
|
||||
}
|
||||
|
||||
export function applyAgentOrchestrationToUserMessage(userMessage, classification, { grantedSkills = [] } = {}) {
|
||||
const displayText = messageDisplayText(userMessage);
|
||||
const skillPrompt = resolveSkillPrompt(classification?.suggestedSkill, grantedSkills);
|
||||
const agentText = buildAgentOrchestrationAgentText({
|
||||
displayText,
|
||||
classification,
|
||||
skillPrompt,
|
||||
});
|
||||
const content = Array.isArray(userMessage?.content)
|
||||
? userMessage.content.map((item, index) => {
|
||||
if (index !== 0) return item;
|
||||
if (typeof item === 'string') return agentText;
|
||||
if (item?.type === 'text') return { ...item, text: agentText };
|
||||
return item;
|
||||
})
|
||||
: [{ type: 'text', text: agentText }];
|
||||
const metadata = {
|
||||
...(userMessage?.metadata ?? {}),
|
||||
displayText: displayText || undefined,
|
||||
chatIntentRoute: CHAT_INTENT_ROUTE.AGENT,
|
||||
chatIntentSource: classification?.source ?? null,
|
||||
};
|
||||
return {
|
||||
...userMessage,
|
||||
content,
|
||||
metadata,
|
||||
};
|
||||
}
|
||||
|
||||
function hasPriorAgentConversation(sessionId, sessionMessageCount) {
|
||||
if (!sessionId || isDirectChatSessionId(sessionId)) return false;
|
||||
if (sessionMessageCount == null) return true;
|
||||
return Number(sessionMessageCount) > 0;
|
||||
}
|
||||
|
||||
export function classifyWithRules({
|
||||
text,
|
||||
forceDeepReasoning = false,
|
||||
toolMode = 'chat',
|
||||
sessionId = null,
|
||||
sessionMessageCount = null,
|
||||
userMessage = null,
|
||||
includeIntentPatterns = true,
|
||||
llmRouterEnabled = false,
|
||||
} = {}) {
|
||||
const normalized = String(text ?? '').trim();
|
||||
if (forceDeepReasoning || toolMode !== 'chat') {
|
||||
return normalizeClassification({
|
||||
route: CHAT_INTENT_ROUTE.AGENT,
|
||||
confidence: 1,
|
||||
reason: forceDeepReasoning ? '用户开启深度推理' : '代码任务模式',
|
||||
}, { source: 'rule' });
|
||||
}
|
||||
if (userMessage && messageHasExplicitChatSkillSelection(userMessage)) {
|
||||
return buildSkillSelectionClassification(userMessage);
|
||||
}
|
||||
if (userMessage && !isTextOnlyUserMessage(userMessage)) {
|
||||
return normalizeClassification({
|
||||
route: CHAT_INTENT_ROUTE.AGENT,
|
||||
confidence: 1,
|
||||
reason: '消息包含非文本内容',
|
||||
}, { source: 'rule' });
|
||||
}
|
||||
if (
|
||||
!llmRouterEnabled &&
|
||||
includeIntentPatterns &&
|
||||
normalized &&
|
||||
isMemoryRecallQuestion(normalized)
|
||||
) {
|
||||
return normalizeClassification({
|
||||
route: CHAT_INTENT_ROUTE.DIRECT_CHAT,
|
||||
confidence: 0.98,
|
||||
reason: '用户在询问个人记忆或历史对话',
|
||||
}, { source: 'rule' });
|
||||
}
|
||||
if (!llmRouterEnabled && hasPriorAgentConversation(sessionId, sessionMessageCount)) {
|
||||
return normalizeClassification({
|
||||
route: CHAT_INTENT_ROUTE.AGENT,
|
||||
confidence: 1,
|
||||
reason: '延续已有 Agent 会话',
|
||||
}, { source: 'rule' });
|
||||
}
|
||||
if (
|
||||
includeIntentPatterns &&
|
||||
normalized &&
|
||||
OBVIOUS_AGENT_PATTERNS.some((pattern) => pattern.test(normalized))
|
||||
) {
|
||||
return normalizeClassification({
|
||||
route: CHAT_INTENT_ROUTE.AGENT,
|
||||
confidence: 0.95,
|
||||
reason: '明确需要生成或发布页面/文件',
|
||||
suggested_skill: 'static-page-publish',
|
||||
agent_brief: '生成或更新 MindSpace 公开页面，并返回可访问链接。',
|
||||
}, { source: 'rule' });
|
||||
}
|
||||
if (
|
||||
includeIntentPatterns &&
|
||||
normalized &&
|
||||
OBVIOUS_DIRECT_PATTERNS.some((pattern) => pattern.test(normalized))
|
||||
) {
|
||||
return normalizeClassification({
|
||||
route: CHAT_INTENT_ROUTE.DIRECT_CHAT,
|
||||
confidence: 0.95,
|
||||
reason: '简单寒暄或连通性测试',
|
||||
}, { source: 'rule' });
|
||||
}
|
||||
return null;
|
||||
}
|
||||
|
||||
export function createChatIntentRouter(options = {}) {
|
||||
const {
|
||||
llmProviderService,
|
||||
memoryV2 = null,
|
||||
env = process.env,
|
||||
logger = console,
|
||||
} = options;
|
||||
const policy = options.policy ?? resolveChatIntentRouterPolicy({
|
||||
env,
|
||||
overrides: pickDefined(options, [
|
||||
'enabled',
|
||||
'modelProviderKeyId',
|
||||
'model',
|
||||
'modelApiType',
|
||||
'temperature',
|
||||
'minConfidence',
|
||||
'memoryResolveEnabled',
|
||||
'memoryResolveLimit',
|
||||
'timeoutMs',
|
||||
'fallbackRoute',
|
||||
]),
|
||||
});
|
||||
|
||||
function getStatus() {
|
||||
return {
|
||||
enabled: Boolean(policy.enabled),
|
||||
modelProviderKeyId: policy.modelProviderKeyId ?? null,
|
||||
model: policy.model ?? null,
|
||||
modelApiType: policy.modelApiType ?? 'chat',
|
||||
minConfidence: policy.minConfidence,
|
||||
memoryResolveEnabled: Boolean(policy.memoryResolveEnabled),
|
||||
memoryResolveLimit: policy.memoryResolveLimit,
|
||||
timeoutMs: policy.timeoutMs,
|
||||
fallbackRoute: policy.fallbackRoute,
|
||||
};
|
||||
}
|
||||
|
||||
function isEnabled() {
|
||||
return Boolean(policy.enabled && llmProviderService?.createChatCompletion);
|
||||
}
|
||||
|
||||
async function resolveRouterContext({ userId, sessionId, text, forceDeepReasoning = false }) {
|
||||
const intervention = resolveMemoryInterventionMode({
|
||||
forceDeepReasoning,
|
||||
recallQuestion: isMemoryRecallQuestion(text),
|
||||
});
|
||||
const limit = memoryLimitForIntervention(intervention, { context: 'router' });
|
||||
if (
|
||||
limit <= 0 ||
|
||||
!policy.memoryResolveEnabled ||
|
||||
!memoryV2?.resolve ||
|
||||
!userId
|
||||
) {
|
||||
return buildRouterContext(null);
|
||||
}
|
||||
try {
|
||||
const resolved = await withTimeout(
|
||||
memoryV2.resolve({
|
||||
userId,
|
||||
sessionId,
|
||||
query: text,
|
||||
limit,
|
||||
}),
|
||||
policy.timeoutMs,
|
||||
'Memory V2 router resolve',
|
||||
);
|
||||
return buildRouterContext(resolved);
|
||||
} catch (err) {
|
||||
logger?.warn?.(
|
||||
`[chat-intent-router] memory resolve skipped: ${err instanceof Error ? err.message : err}`,
|
||||
);
|
||||
return buildRouterContext({
|
||||
degraded: true,
|
||||
reason: 'memory_resolve_failed',
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
async function classify({
|
||||
userId = null,
|
||||
userMessage,
|
||||
sessionId = null,
|
||||
sessionMessageCount = null,
|
||||
toolMode = 'chat',
|
||||
forceDeepReasoning = false,
|
||||
grantedSkills = [],
|
||||
} = {}) {
|
||||
const text = messageDisplayText(userMessage);
|
||||
const ruleResult = classifyWithRules({
|
||||
text,
|
||||
forceDeepReasoning,
|
||||
toolMode,
|
||||
sessionId,
|
||||
sessionMessageCount,
|
||||
userMessage,
|
||||
includeIntentPatterns: false,
|
||||
llmRouterEnabled: Boolean(policy.enabled),
|
||||
});
|
||||
if (ruleResult) return ruleResult;
|
||||
if (!isEnabled()) {
|
||||
return normalizeClassification({
|
||||
route: policy.fallbackRoute,
|
||||
confidence: 0.5,
|
||||
reason: '意图路由未启用，走默认通道',
|
||||
}, { source: 'fallback' });
|
||||
}
|
||||
|
||||
const routerContext = await resolveRouterContext({
|
||||
userId,
|
||||
sessionId,
|
||||
text,
|
||||
forceDeepReasoning,
|
||||
});
|
||||
let completion = null;
|
||||
try {
|
||||
completion = await withTimeout(
|
||||
llmProviderService.createChatCompletion({
|
||||
providerKeyId: policy.modelProviderKeyId || undefined,
|
||||
model: policy.model || undefined,
|
||||
modelApiType: policy.modelApiType || undefined,
|
||||
temperature: policy.temperature,
|
||||
messages: [
|
||||
{ role: 'system', content: buildRouterSystemPrompt(grantedSkills) },
|
||||
{
|
||||
role: 'user',
|
||||
content: buildRouterUserPrompt({
|
||||
text,
|
||||
routerContext: routerContext.content,
|
||||
}),
|
||||
},
|
||||
],
|
||||
}),
|
||||
policy.timeoutMs,
|
||||
'Chat intent router',
|
||||
);
|
||||
} catch (err) {
|
||||
return normalizeClassification({
|
||||
route: policy.fallbackRoute,
|
||||
confidence: 0,
|
||||
reason: err instanceof Error ? err.message : '意图路由失败，走默认通道',
|
||||
memory: routerContext,
|
||||
}, { source: 'fallback', fallbackRoute: policy.fallbackRoute });
|
||||
}
|
||||
if (!completion?.ok) {
|
||||
return normalizeClassification({
|
||||
route: policy.fallbackRoute,
|
||||
confidence: 0,
|
||||
reason: completion?.message ?? '意图路由失败，走默认通道',
|
||||
memory: routerContext,
|
||||
}, { source: 'fallback', fallbackRoute: policy.fallbackRoute });
|
||||
}
|
||||
|
||||
const parsed = parseRouterJson(completion.reply);
|
||||
if (!parsed) {
|
||||
return normalizeClassification({
|
||||
route: policy.fallbackRoute,
|
||||
confidence: 0,
|
||||
reason: '意图路由响应无法解析，走默认通道',
|
||||
memory: routerContext,
|
||||
}, { source: 'fallback', fallbackRoute: policy.fallbackRoute });
|
||||
}
|
||||
|
||||
const classification = {
|
||||
...normalizeClassification(parsed, { source: 'llm', fallbackRoute: policy.fallbackRoute }),
|
||||
providerKeyId: completion.providerKeyId ?? policy.modelProviderKeyId ?? null,
|
||||
model: completion.model ?? policy.model ?? null,
|
||||
memory: routerContext,
|
||||
};
|
||||
if (
|
||||
classification.route === CHAT_INTENT_ROUTE.DIRECT_CHAT &&
|
||||
classification.confidence < policy.minConfidence
|
||||
) {
|
||||
return normalizeClassification({
|
||||
...classification,
|
||||
route: policy.fallbackRoute,
|
||||
reason: `${classification.reason}（置信度 ${classification.confidence} 低于阈值，走默认通道）`,
|
||||
}, { source: 'threshold', fallbackRoute: policy.fallbackRoute });
|
||||
}
|
||||
return classification;
|
||||
}
|
||||
|
||||
return {
|
||||
getStatus,
|
||||
isEnabled,
|
||||
classify,
|
||||
applyAgentOrchestration: applyAgentOrchestrationToUserMessage,
|
||||
};
|
||||
}
|
||||
|
||||
export function createManagedChatIntentRouter({
|
||||
llmProviderService,
|
||||
memoryV2 = null,
|
||||
configService = null,
|
||||
env = process.env,
|
||||
logger = console,
|
||||
} = {}) {
|
||||
let activeRouter = null;
|
||||
let activeFingerprint = null;
|
||||
let activeMeta = { source: 'env', updatedAt: null, updatedBy: null, configError: null };
|
||||
|
||||
async function loadRouterState() {
|
||||
if (!configService?.getRuntimeState) {
|
||||
return {
|
||||
source: 'env',
|
||||
updatedAt: null,
|
||||
updatedBy: null,
|
||||
fingerprint: 'env-only',
|
||||
effectiveEnv: { ...env },
|
||||
configError: null,
|
||||
};
|
||||
}
|
||||
try {
|
||||
const state = await configService.getRuntimeState();
|
||||
return {
|
||||
source: state?.source ?? 'admin-db',
|
||||
updatedAt: state?.updatedAt ?? null,
|
||||
updatedBy: state?.updatedBy ?? null,
|
||||
fingerprint: state?.fingerprint ?? `admin-db:${Date.now()}`,
|
||||
effectiveEnv: { ...env, ...(state?.overrides ?? {}) },
|
||||
configError: null,
|
||||
};
|
||||
} catch (err) {
|
||||
logger?.warn?.(
|
||||
`[chat-intent-router] admin config unavailable, using process env: ${err instanceof Error ? err.message : err}`,
|
||||
);
|
||||
return {
|
||||
source: 'env-fallback',
|
||||
updatedAt: null,
|
||||
updatedBy: null,
|
||||
fingerprint: 'env-fallback',
|
||||
effectiveEnv: { ...env },
|
||||
configError: err instanceof Error ? err.message : String(err),
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
async function ensureRouter() {
|
||||
const state = await loadRouterState();
|
||||
if (activeRouter && activeFingerprint === state.fingerprint) {
|
||||
activeMeta = state;
|
||||
return activeRouter;
|
||||
}
|
||||
activeRouter = createChatIntentRouter({
|
||||
llmProviderService,
|
||||
memoryV2,
|
||||
env: state.effectiveEnv,
|
||||
logger,
|
||||
});
|
||||
activeFingerprint = state.fingerprint;
|
||||
activeMeta = state;
|
||||
return activeRouter;
|
||||
}
|
||||
|
||||
return {
|
||||
async getStatus() {
|
||||
const router = await ensureRouter();
|
||||
return {
|
||||
...router.getStatus(),
|
||||
configSource: activeMeta.source,
|
||||
configUpdatedAt: activeMeta.updatedAt,
|
||||
configUpdatedBy: activeMeta.updatedBy,
|
||||
configError: activeMeta.configError,
|
||||
};
|
||||
},
|
||||
|
||||
async isEnabled() {
|
||||
const router = await ensureRouter();
|
||||
return router.isEnabled();
|
||||
},
|
||||
|
||||
async classify(input = {}) {
|
||||
const router = await ensureRouter();
|
||||
return router.classify(input);
|
||||
},
|
||||
|
||||
applyAgentOrchestration: applyAgentOrchestrationToUserMessage,
|
||||
};
|
||||
}
|
||||
Reference in New Issue
Block a user