15f84bf8c1
Some checks failed
CI / Lint & TypeCheck (push) Has been cancelled
CI / Unit Tests (push) Has been cancelled
CI / Build Frontend (push) Has been cancelled
CI / Rust Check (push) Has been cancelled
CI / Security Scan (push) Has been cancelled
CI / E2E Tests (push) Has been cancelled
suggestion prompt 新增规则：建议会被用户直接点击发送， 因此不包含"领导/老板/老师"等称谓，改用无主语句式。 同步更新示例和关怀模板中的表达方式。
917 lines
27 KiB
TypeScript
917 lines
27 KiB
TypeScript
/**
|
||
* LLM Service Adapter - Unified LLM interface for L4 self-evolution engines
|
||
*
|
||
* Provides a unified interface for:
|
||
* - ReflectionEngine: Semantic analysis + deep reflection
|
||
* - ContextCompactor: High-quality summarization
|
||
* - MemoryExtractor: Semantic importance scoring
|
||
*
|
||
* Supports multiple backends:
|
||
* - OpenAI (GPT-4, GPT-3.5)
|
||
* - Volcengine (Doubao)
|
||
* - ZCLAW Gateway (passthrough)
|
||
*
|
||
* Part of ZCLAW L4 Self-Evolution capability.
|
||
*/
|
||
|
||
import { DEFAULT_MODEL_ID, DEFAULT_OPENAI_BASE_URL } from '../constants/models';
|
||
import { createLogger } from './logger';
|
||
import { recordLLMUsage } from './telemetry-collector';
|
||
|
||
const log = createLogger('LLMService');
|
||
|
||
// === Types ===
|
||
|
||
export type LLMProvider = 'openai' | 'volcengine' | 'gateway' | 'saas' | 'mock';
|
||
|
||
export interface LLMConfig {
|
||
provider: LLMProvider;
|
||
model?: string;
|
||
apiKey?: string;
|
||
apiBase?: string;
|
||
maxTokens?: number;
|
||
temperature?: number;
|
||
timeout?: number;
|
||
}
|
||
|
||
export interface LLMMessage {
|
||
role: 'system' | 'user' | 'assistant';
|
||
content: string;
|
||
}
|
||
|
||
export interface LLMResponse {
|
||
content: string;
|
||
tokensUsed?: {
|
||
input: number;
|
||
output: number;
|
||
};
|
||
model?: string;
|
||
latencyMs?: number;
|
||
}
|
||
|
||
export interface LLMServiceAdapter {
|
||
complete(messages: LLMMessage[], options?: Partial<LLMConfig>): Promise<LLMResponse>;
|
||
isAvailable(): boolean;
|
||
getProvider(): LLMProvider;
|
||
}
|
||
|
||
// === Default Configs ===
|
||
|
||
const DEFAULT_CONFIGS: Record<LLMProvider, LLMConfig> = {
|
||
openai: {
|
||
provider: 'openai',
|
||
model: DEFAULT_MODEL_ID,
|
||
apiBase: DEFAULT_OPENAI_BASE_URL,
|
||
maxTokens: 2000,
|
||
temperature: 0.7,
|
||
timeout: 30000,
|
||
},
|
||
volcengine: {
|
||
provider: 'volcengine',
|
||
model: 'doubao-pro-32k',
|
||
apiBase: 'https://ark.cn-beijing.volces.com/api/v3',
|
||
maxTokens: 2000,
|
||
temperature: 0.7,
|
||
timeout: 30000,
|
||
},
|
||
gateway: {
|
||
provider: 'gateway',
|
||
apiBase: '/api/llm',
|
||
maxTokens: 2000,
|
||
temperature: 0.7,
|
||
timeout: 60000,
|
||
},
|
||
saas: {
|
||
provider: 'saas',
|
||
maxTokens: 4096,
|
||
temperature: 0.7,
|
||
timeout: 300000, // 5 min for streaming
|
||
},
|
||
mock: {
|
||
provider: 'mock',
|
||
maxTokens: 100,
|
||
temperature: 0,
|
||
timeout: 100,
|
||
},
|
||
};
|
||
|
||
// === Storage ===
|
||
|
||
const LLM_CONFIG_KEY = 'zclaw-llm-config';
|
||
|
||
// === Mock Adapter (for testing) ===
|
||
|
||
class MockLLMAdapter implements LLMServiceAdapter {
|
||
constructor(_config: LLMConfig) {
|
||
// Config is stored for future use (e.g., custom mock behavior based on config)
|
||
}
|
||
|
||
async complete(messages: LLMMessage[]): Promise<LLMResponse> {
|
||
// Simulate latency
|
||
await new Promise((resolve) => setTimeout(resolve, 50));
|
||
|
||
const lastMessage = messages[messages.length - 1];
|
||
const content = lastMessage?.content || '';
|
||
|
||
// Generate mock response based on content type
|
||
let response = '[Mock LLM Response] ';
|
||
|
||
if (content.includes('reflect') || content.includes('反思')) {
|
||
response += JSON.stringify({
|
||
patterns: [
|
||
{
|
||
observation: '用户经常询问代码优化相关问题',
|
||
frequency: 5,
|
||
sentiment: 'positive',
|
||
evidence: ['多次讨论性能优化', '关注代码质量'],
|
||
},
|
||
],
|
||
improvements: [
|
||
{
|
||
area: '代码解释',
|
||
suggestion: '可以提供更详细的代码注释',
|
||
priority: 'medium',
|
||
},
|
||
],
|
||
identityProposals: [],
|
||
});
|
||
} else if (content.includes('summarize') || content.includes('摘要')) {
|
||
response += '这是一个关于对话内容的摘要，包含了主要讨论的要点和结论。';
|
||
} else if (content.includes('importance') || content.includes('重要性')) {
|
||
response += JSON.stringify({
|
||
memories: [
|
||
{ content: '用户偏好简洁的回答', importance: 7, type: 'preference' },
|
||
],
|
||
});
|
||
} else {
|
||
response += 'Processed: ' + content.slice(0, 50);
|
||
}
|
||
|
||
return {
|
||
content: response,
|
||
tokensUsed: { input: content.length / 4, output: response.length / 4 },
|
||
model: 'mock-model',
|
||
latencyMs: 50,
|
||
};
|
||
}
|
||
|
||
isAvailable(): boolean {
|
||
return true;
|
||
}
|
||
|
||
getProvider(): LLMProvider {
|
||
return 'mock';
|
||
}
|
||
}
|
||
|
||
// === OpenAI Adapter ===
|
||
|
||
class OpenAILLMAdapter implements LLMServiceAdapter {
|
||
private config: LLMConfig;
|
||
|
||
constructor(config: LLMConfig) {
|
||
this.config = { ...DEFAULT_CONFIGS.openai, ...config };
|
||
}
|
||
|
||
async complete(messages: LLMMessage[], options?: Partial<LLMConfig>): Promise<LLMResponse> {
|
||
const config = { ...this.config, ...options };
|
||
const startTime = Date.now();
|
||
|
||
if (!config.apiKey) {
|
||
throw new Error('[OpenAI] API key not configured');
|
||
}
|
||
|
||
const response = await fetch(`${config.apiBase}/chat/completions`, {
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${config.apiKey}`,
|
||
},
|
||
body: JSON.stringify({
|
||
model: config.model,
|
||
messages,
|
||
max_tokens: config.maxTokens,
|
||
temperature: config.temperature,
|
||
}),
|
||
signal: AbortSignal.timeout(config.timeout || 30000),
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const errorBody = await response.text();
|
||
// Log full error in development only
|
||
if (import.meta.env.DEV) {
|
||
console.error('[OpenAI] API error:', errorBody);
|
||
}
|
||
// Return sanitized error to caller
|
||
throw new Error(`[OpenAI] API error: ${response.status} - Request failed`);
|
||
}
|
||
|
||
const data = await response.json();
|
||
const latencyMs = Date.now() - startTime;
|
||
|
||
return {
|
||
content: data.choices[0]?.message?.content || '',
|
||
tokensUsed: {
|
||
input: data.usage?.prompt_tokens || 0,
|
||
output: data.usage?.completion_tokens || 0,
|
||
},
|
||
model: data.model,
|
||
latencyMs,
|
||
};
|
||
}
|
||
|
||
isAvailable(): boolean {
|
||
return !!this.config.apiKey;
|
||
}
|
||
|
||
getProvider(): LLMProvider {
|
||
return 'openai';
|
||
}
|
||
}
|
||
|
||
// === Volcengine Adapter ===
|
||
|
||
class VolcengineLLMAdapter implements LLMServiceAdapter {
|
||
private config: LLMConfig;
|
||
|
||
constructor(config: LLMConfig) {
|
||
this.config = { ...DEFAULT_CONFIGS.volcengine, ...config };
|
||
}
|
||
|
||
async complete(messages: LLMMessage[], options?: Partial<LLMConfig>): Promise<LLMResponse> {
|
||
const config = { ...this.config, ...options };
|
||
const startTime = Date.now();
|
||
|
||
if (!config.apiKey) {
|
||
throw new Error('[Volcengine] API key not configured');
|
||
}
|
||
|
||
const response = await fetch(`${config.apiBase}/chat/completions`, {
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${config.apiKey}`,
|
||
},
|
||
body: JSON.stringify({
|
||
model: config.model,
|
||
messages,
|
||
max_tokens: config.maxTokens,
|
||
temperature: config.temperature,
|
||
}),
|
||
signal: AbortSignal.timeout(config.timeout || 30000),
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const errorBody = await response.text();
|
||
// Log full error in development only
|
||
if (import.meta.env.DEV) {
|
||
console.error('[Volcengine] API error:', errorBody);
|
||
}
|
||
// Return sanitized error to caller
|
||
throw new Error(`[Volcengine] API error: ${response.status} - Request failed`);
|
||
}
|
||
|
||
const data = await response.json();
|
||
const latencyMs = Date.now() - startTime;
|
||
|
||
return {
|
||
content: data.choices[0]?.message?.content || '',
|
||
tokensUsed: {
|
||
input: data.usage?.prompt_tokens || 0,
|
||
output: data.usage?.completion_tokens || 0,
|
||
},
|
||
model: data.model,
|
||
latencyMs,
|
||
};
|
||
}
|
||
|
||
isAvailable(): boolean {
|
||
return !!this.config.apiKey;
|
||
}
|
||
|
||
getProvider(): LLMProvider {
|
||
return 'volcengine';
|
||
}
|
||
}
|
||
|
||
// === Gateway Adapter (pass through to ZCLAW or internal Kernel) ===
|
||
|
||
class GatewayLLMAdapter implements LLMServiceAdapter {
|
||
private config: LLMConfig;
|
||
|
||
constructor(config: LLMConfig) {
|
||
this.config = { ...DEFAULT_CONFIGS.gateway, ...config };
|
||
}
|
||
|
||
async complete(messages: LLMMessage[], options?: Partial<LLMConfig>): Promise<LLMResponse> {
|
||
const config = { ...this.config, ...options };
|
||
const startTime = Date.now();
|
||
|
||
// Build a single prompt from messages
|
||
const systemMessage = messages.find(m => m.role === 'system')?.content || '';
|
||
const userMessage = messages.find(m => m.role === 'user')?.content || '';
|
||
|
||
// Combine system and user messages into a single prompt
|
||
const fullPrompt = systemMessage
|
||
? `${systemMessage}\n\n${userMessage}`
|
||
: userMessage;
|
||
|
||
// Check if running in Tauri with internal kernel
|
||
// Use the same detection as kernel-client.ts
|
||
const isTauri = typeof window !== 'undefined' &&
|
||
'__TAURI_INTERNALS__' in window;
|
||
|
||
if (isTauri) {
|
||
// Use internal Kernel via Tauri invoke
|
||
try {
|
||
const { invoke } = await import('@tauri-apps/api/core');
|
||
|
||
// Get the default agent ID from connectionStore or use the first agent
|
||
const agentId = localStorage.getItem('zclaw-default-agent-id');
|
||
|
||
const response = await invoke<{ content: string; input_tokens: number; output_tokens: number }>('agent_chat', {
|
||
request: {
|
||
agentId: agentId || null, // null will use default agent
|
||
message: fullPrompt,
|
||
},
|
||
});
|
||
|
||
const latencyMs = Date.now() - startTime;
|
||
return {
|
||
content: response.content || '',
|
||
tokensUsed: {
|
||
input: response.input_tokens || 0,
|
||
output: response.output_tokens || 0,
|
||
},
|
||
latencyMs,
|
||
};
|
||
} catch (err) {
|
||
console.error('[LLMService] Kernel chat failed:', err);
|
||
const message = err instanceof Error ? err.message : String(err);
|
||
throw new Error(`[Gateway] Kernel chat failed: ${message}`);
|
||
}
|
||
}
|
||
|
||
// External Gateway mode: Use ZCLAW's chat endpoint
|
||
const agentId = localStorage.getItem('zclaw-default-agent-id') || 'default';
|
||
|
||
const response = await fetch(`/api/agents/${agentId}/message`, {
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
},
|
||
body: JSON.stringify({
|
||
message: fullPrompt,
|
||
max_tokens: config.maxTokens,
|
||
temperature: config.temperature ?? 0.3, // Lower temperature for extraction tasks
|
||
}),
|
||
signal: AbortSignal.timeout(config.timeout || 60000),
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const error = await response.text();
|
||
// If agent not found, try without agent ID (direct /api/chat)
|
||
if (response.status === 404) {
|
||
const fallbackResponse = await fetch('/api/chat', {
|
||
method: 'POST',
|
||
headers: { 'Content-Type': 'application/json' },
|
||
body: JSON.stringify({
|
||
message: fullPrompt,
|
||
max_tokens: config.maxTokens,
|
||
temperature: config.temperature ?? 0.3,
|
||
}),
|
||
signal: AbortSignal.timeout(config.timeout || 60000),
|
||
});
|
||
|
||
if (!fallbackResponse.ok) {
|
||
throw new Error(`[Gateway] Both endpoints failed: ${fallbackResponse.status}`);
|
||
}
|
||
|
||
const data = await fallbackResponse.json();
|
||
const latencyMs = Date.now() - startTime;
|
||
return {
|
||
content: data.response || data.content || '',
|
||
tokensUsed: { input: data.input_tokens || 0, output: data.output_tokens || 0 },
|
||
latencyMs,
|
||
};
|
||
}
|
||
throw new Error(`[Gateway] API error: ${response.status} - ${error}`);
|
||
}
|
||
|
||
const data = await response.json();
|
||
const latencyMs = Date.now() - startTime;
|
||
|
||
return {
|
||
content: data.response || data.content || '',
|
||
tokensUsed: { input: data.input_tokens || 0, output: data.output_tokens || 0 },
|
||
latencyMs,
|
||
};
|
||
}
|
||
|
||
isAvailable(): boolean {
|
||
// Gateway is available if we're in browser (can connect to ZCLAW)
|
||
return typeof window !== 'undefined';
|
||
}
|
||
|
||
getProvider(): LLMProvider {
|
||
return 'gateway';
|
||
}
|
||
}
|
||
|
||
// === SaaS Relay Adapter (via SaaS backend) ===
|
||
|
||
class SaasLLMAdapter implements LLMServiceAdapter {
|
||
private config: LLMConfig;
|
||
|
||
constructor(config: LLMConfig) {
|
||
this.config = { ...DEFAULT_CONFIGS.saas, ...config };
|
||
}
|
||
|
||
async complete(messages: LLMMessage[], options?: Partial<LLMConfig>): Promise<LLMResponse> {
|
||
const config = { ...this.config, ...options };
|
||
const startTime = Date.now();
|
||
|
||
// Dynamic import to avoid circular dependency
|
||
const { useSaaSStore } = await import('../store/saasStore');
|
||
const { saasUrl, authToken } = useSaaSStore.getState();
|
||
|
||
if (!saasUrl || !authToken) {
|
||
throw new Error('[SaaS] 未登录 SaaS 平台，请先在设置中登录');
|
||
}
|
||
|
||
// Dynamic import of SaaSClient singleton
|
||
const { saasClient } = await import('./saas-client');
|
||
saasClient.setBaseUrl(saasUrl);
|
||
saasClient.setToken(authToken);
|
||
|
||
const openaiBody = {
|
||
model: config.model || 'default',
|
||
messages,
|
||
max_tokens: config.maxTokens || 4096,
|
||
temperature: config.temperature ?? 0.7,
|
||
stream: false,
|
||
};
|
||
|
||
const response = await saasClient.chatCompletion(
|
||
openaiBody,
|
||
AbortSignal.timeout(config.timeout || 300000),
|
||
);
|
||
|
||
if (!response.ok) {
|
||
const errorData = await response.json().catch(() => ({
|
||
error: 'unknown',
|
||
message: `SaaS relay 请求失败 (${response.status})`,
|
||
}));
|
||
throw new Error(
|
||
`[SaaS] ${errorData.message || errorData.error || `请求失败: ${response.status}`}`,
|
||
);
|
||
}
|
||
|
||
const data = await response.json();
|
||
const latencyMs = Date.now() - startTime;
|
||
|
||
const result = {
|
||
content: data.choices?.[0]?.message?.content || '',
|
||
tokensUsed: {
|
||
input: data.usage?.prompt_tokens || 0,
|
||
output: data.usage?.completion_tokens || 0,
|
||
},
|
||
model: data.model,
|
||
latencyMs,
|
||
};
|
||
|
||
// Record telemetry for SaaS relay usage
|
||
try {
|
||
const { recordLLMUsage } = await import('./telemetry-collector');
|
||
recordLLMUsage(
|
||
result.model || 'saas-relay',
|
||
result.tokensUsed.input,
|
||
result.tokensUsed.output,
|
||
{ latencyMs, success: true, connectionMode: 'saas' },
|
||
);
|
||
} catch (e) {
|
||
log.debug('Failed to record LLM telemetry', { error: e });
|
||
}
|
||
|
||
return result;
|
||
}
|
||
|
||
isAvailable(): boolean {
|
||
// Check synchronously via localStorage for availability check.
|
||
// Auth is cookie-based — check connection mode + URL presence.
|
||
try {
|
||
const mode = localStorage.getItem('zclaw-connection-mode');
|
||
const saasUrl = localStorage.getItem('zclaw-saas-url');
|
||
return mode === 'saas' && !!saasUrl;
|
||
} catch (e) {
|
||
log.debug('Failed to check SaaS adapter availability', { error: e });
|
||
return false;
|
||
}
|
||
}
|
||
|
||
getProvider(): LLMProvider {
|
||
return 'saas';
|
||
}
|
||
}
|
||
|
||
// === Factory ===
|
||
|
||
let cachedAdapter: LLMServiceAdapter | null = null;
|
||
|
||
export function createLLMAdapter(config?: Partial<LLMConfig>): LLMServiceAdapter {
|
||
const savedConfig = loadConfig();
|
||
const finalConfig = { ...savedConfig, ...config };
|
||
|
||
switch (finalConfig.provider) {
|
||
case 'openai':
|
||
return new OpenAILLMAdapter(finalConfig);
|
||
case 'volcengine':
|
||
return new VolcengineLLMAdapter(finalConfig);
|
||
case 'gateway':
|
||
return new GatewayLLMAdapter(finalConfig);
|
||
case 'saas':
|
||
return new SaasLLMAdapter(finalConfig);
|
||
case 'mock':
|
||
default:
|
||
return new MockLLMAdapter(finalConfig);
|
||
}
|
||
}
|
||
|
||
export function getLLMAdapter(): LLMServiceAdapter {
|
||
if (!cachedAdapter) {
|
||
cachedAdapter = createLLMAdapter();
|
||
}
|
||
return cachedAdapter;
|
||
}
|
||
|
||
export function resetLLMAdapter(): void {
|
||
cachedAdapter = null;
|
||
}
|
||
|
||
// === Config Management ===
|
||
|
||
export function loadConfig(): LLMConfig {
|
||
if (typeof window === 'undefined') {
|
||
return DEFAULT_CONFIGS.mock;
|
||
}
|
||
|
||
try {
|
||
const saved = localStorage.getItem(LLM_CONFIG_KEY);
|
||
if (saved) {
|
||
return JSON.parse(saved);
|
||
}
|
||
} catch (e) {
|
||
log.debug('Failed to parse LLM config', { error: e });
|
||
}
|
||
|
||
// Default to gateway (ZCLAW passthrough) for L4 self-evolution
|
||
return DEFAULT_CONFIGS.gateway;
|
||
}
|
||
|
||
export function saveConfig(config: LLMConfig): void {
|
||
if (typeof window === 'undefined') return;
|
||
|
||
// Don't save API key to localStorage for security
|
||
const safeConfig = { ...config };
|
||
delete safeConfig.apiKey;
|
||
|
||
localStorage.setItem(LLM_CONFIG_KEY, JSON.stringify(safeConfig));
|
||
|
||
// Mark config as dirty for SaaS push sync
|
||
localStorage.setItem('zclaw-config-dirty.llm.default', '1');
|
||
|
||
resetLLMAdapter();
|
||
}
|
||
|
||
// === Prompt Templates ===
|
||
|
||
// 硬编码默认值 — 当 SaaS 不可用且本地无缓存时的终极兜底
|
||
const HARDCODED_PROMPTS: Record<string, { system: string; user: (arg: string) => string }> = {
|
||
reflection: {
|
||
system: `你是一个 AI Agent 的自我反思引擎。分析最近的对话历史，识别行为模式，并生成改进建议。
|
||
|
||
输出 JSON 格式：
|
||
{
|
||
"patterns": [
|
||
{
|
||
"observation": "观察到的模式描述",
|
||
"frequency": 数字,
|
||
"sentiment": "positive/negative/neutral",
|
||
"evidence": ["证据1", "证据2"]
|
||
}
|
||
],
|
||
"improvements": [
|
||
{
|
||
"area": "改进领域",
|
||
"suggestion": "具体建议",
|
||
"priority": "high/medium/low"
|
||
}
|
||
],
|
||
"identityProposals": []
|
||
}`,
|
||
user: (context: string) => `分析以下对话历史，进行自我反思：\n\n${context}\n\n请识别行为模式（积极和消极），并提供具体的改进建议。`,
|
||
},
|
||
|
||
compaction: {
|
||
system: `你是一个对话摘要专家。将长对话压缩为简洁的摘要，保留关键信息。
|
||
|
||
要求：
|
||
1. 保留所有重要决策和结论
|
||
2. 保留用户偏好和约束
|
||
3. 保留未完成的任务
|
||
4. 保持时间顺序
|
||
5. 摘要应能在后续对话中替代原始内容`,
|
||
user: (messages: string) => `请将以下对话压缩为简洁摘要，保留关键信息：\n\n${messages}`,
|
||
},
|
||
|
||
extraction: {
|
||
system: `你是一个记忆提取专家。从对话中提取值得长期记住的信息。
|
||
|
||
提取类型：
|
||
- fact: 用户告知的事实（如"我的公司叫XXX"）
|
||
- preference: 用户的偏好（如"我喜欢简洁的回答"）
|
||
- lesson: 本次对话的经验教训
|
||
- task: 未完成的任务或承诺
|
||
|
||
输出 JSON 数组：
|
||
[
|
||
{
|
||
"content": "记忆内容",
|
||
"type": "fact/preference/lesson/task",
|
||
"importance": 1-10,
|
||
"tags": ["标签1", "标签2"]
|
||
}
|
||
]`,
|
||
user: (conversation: string) => `从以下对话中提取值得长期记住的信息：\n\n${conversation}\n\n如果没有值得记忆的内容，返回空数组 []。`,
|
||
},
|
||
|
||
suggestions: {
|
||
system: `你是 ZCLAW 的管家助手，需要站在用户角度思考他们真正需要什么，生成 3 个个性化建议。
|
||
|
||
## 生成规则
|
||
1. 第 1 条 — 深入追问：基于当前话题，提出一个有洞察力的追问，帮助用户深入探索
|
||
2. 第 2 条 — 实用行动：建议一个具体的、可操作的下一步（调用技能、执行工具、查看数据等）
|
||
3. 第 3 条 — 管家关怀：
|
||
- 如果有未解决痛点 → 回访建议，如"上次提到的X，后来解决了吗？"
|
||
- 如果有相关经验 → 引导复用，如"上次用X方法解决了类似问题，要再试试吗？"
|
||
- 如果有匹配技能 → 推荐使用，如"试试 [技能名] 来处理这个"
|
||
- 如果没有提供痛点/经验/技能信息 → 给出一个启发性的思考角度
|
||
4. 每个不超过 30 个中文字符
|
||
5. 不要重复对话中已讨论过的内容
|
||
6. 不要生成空泛的建议（如"继续分析"、"换个角度"）
|
||
7. 使用与用户相同的语言
|
||
8. 建议会被用户直接点击发送，因此不要包含任何称谓（如"领导"、"老板"、"老师"等），用无主语的问句或陈述句
|
||
|
||
只输出 JSON 数组，包含恰好 3 个字符串。不要输出任何其他内容。
|
||
示例：["科室绩效分析可以按哪些维度拆解？", "用 researcher 技能查一下相关文献？", "上次提到的排班冲突问题，需要继续想解决方案吗？"]`,
|
||
user: (context: string) => `以下是对话中最近的消息：\n\n${context}\n\n请生成 3 个后续建议（1 深入追问 + 1 实用行动 + 1 管家关怀）。`,
|
||
},
|
||
};
|
||
|
||
// === Prompt Cache (SaaS OTA) ===
|
||
|
||
const PROMPT_CACHE_KEY = 'zclaw-prompt-cache';
|
||
|
||
interface CachedPrompt {
|
||
name: string;
|
||
version: number;
|
||
source: string;
|
||
system: string;
|
||
userTemplate: string | null;
|
||
syncedAt: string;
|
||
}
|
||
|
||
/** 获取本地缓存的提示词版本号映射 */
|
||
function loadPromptCache(): Record<string, CachedPrompt> {
|
||
if (typeof window === 'undefined') return {};
|
||
try {
|
||
const raw = localStorage.getItem(PROMPT_CACHE_KEY);
|
||
return raw ? JSON.parse(raw) : {};
|
||
} catch (e) {
|
||
log.debug('Failed to parse prompt cache', { error: e });
|
||
return {};
|
||
}
|
||
}
|
||
|
||
/** 保存提示词缓存到 localStorage */
|
||
function savePromptCache(cache: Record<string, CachedPrompt>): void {
|
||
if (typeof window === 'undefined') return;
|
||
localStorage.setItem(PROMPT_CACHE_KEY, JSON.stringify(cache));
|
||
}
|
||
|
||
/**
|
||
* 获取指定提示词的系统提示词
|
||
* 优先级：本地缓存 → 硬编码默认值
|
||
*/
|
||
export function getSystemPrompt(name: string): string {
|
||
const cache = loadPromptCache();
|
||
if (cache[name]?.system) {
|
||
return cache[name].system;
|
||
}
|
||
return HARDCODED_PROMPTS[name]?.system ?? '';
|
||
}
|
||
|
||
/**
|
||
* 获取指定提示词的用户提示词模板
|
||
* 优先级：本地缓存 → 硬编码默认值
|
||
*/
|
||
export function getUserPromptTemplate(name: string): string | ((arg: string) => string) | null {
|
||
const cache = loadPromptCache();
|
||
if (cache[name]) {
|
||
const tmpl = cache[name].userTemplate;
|
||
if (tmpl) return tmpl;
|
||
}
|
||
return HARDCODED_PROMPTS[name]?.user ?? null;
|
||
}
|
||
|
||
/** 获取提示词当前版本号（本地缓存） */
|
||
export function getPromptVersion(name: string): number {
|
||
const cache = loadPromptCache();
|
||
return cache[name]?.version ?? 0;
|
||
}
|
||
|
||
/** 获取所有本地缓存的提示词版本（用于 OTA 检查） */
|
||
export function getAllPromptVersions(): Record<string, number> {
|
||
const cache = loadPromptCache();
|
||
const versions: Record<string, number> = {};
|
||
for (const [name, entry] of Object.entries(cache)) {
|
||
versions[name] = entry.version;
|
||
}
|
||
return versions;
|
||
}
|
||
|
||
/**
|
||
* 应用 SaaS OTA 更新到本地缓存
|
||
* @param updates 从 SaaS 拉取的更新列表
|
||
*/
|
||
export function applyPromptUpdates(updates: Array<{
|
||
name: string;
|
||
version: number;
|
||
system_prompt: string;
|
||
user_prompt_template: string | null;
|
||
source: string;
|
||
changelog?: string | null;
|
||
}>): number {
|
||
const cache = loadPromptCache();
|
||
let applied = 0;
|
||
|
||
for (const update of updates) {
|
||
cache[update.name] = {
|
||
name: update.name,
|
||
version: update.version,
|
||
source: update.source,
|
||
system: update.system_prompt,
|
||
userTemplate: update.user_prompt_template,
|
||
syncedAt: new Date().toISOString(),
|
||
};
|
||
applied++;
|
||
}
|
||
|
||
if (applied > 0) {
|
||
savePromptCache(cache);
|
||
}
|
||
|
||
return applied;
|
||
}
|
||
|
||
/**
|
||
* 后台异步检查 SaaS 提示词更新
|
||
* 启动时和每 30 分钟调用一次
|
||
*/
|
||
let promptSyncTimer: ReturnType<typeof setInterval> | null = null;
|
||
|
||
export function startPromptOTASync(deviceId: string): void {
|
||
if (promptSyncTimer) return; // 已启动
|
||
if (typeof window === 'undefined') return;
|
||
|
||
const doSync = async () => {
|
||
try {
|
||
const { saasClient } = await import('./saas-client');
|
||
const { useSaaSStore } = await import('../store/saasStore');
|
||
const { saasUrl, authToken } = useSaaSStore.getState();
|
||
|
||
if (!saasUrl || !authToken) return;
|
||
|
||
saasClient.setBaseUrl(saasUrl);
|
||
saasClient.setToken(authToken);
|
||
|
||
const versions = getAllPromptVersions();
|
||
const result = await saasClient.checkPromptUpdates(deviceId, versions);
|
||
|
||
if (result.updates.length > 0) {
|
||
const applied = applyPromptUpdates(result.updates);
|
||
if (applied > 0) {
|
||
log.debug(`已更新 ${applied} 个提示词模板`);
|
||
}
|
||
}
|
||
} catch (err) {
|
||
// 静默失败，不影响正常使用
|
||
log.debug('检查更新失败:', err);
|
||
}
|
||
};
|
||
|
||
// 立即执行一次
|
||
doSync();
|
||
// 每 30 分钟检查一次
|
||
promptSyncTimer = setInterval(doSync, 30 * 60 * 1000);
|
||
}
|
||
|
||
export function stopPromptOTASync(): void {
|
||
if (promptSyncTimer) {
|
||
clearInterval(promptSyncTimer);
|
||
promptSyncTimer = null;
|
||
}
|
||
}
|
||
|
||
// 保留向后兼容的 LLM_PROMPTS 导出（读取走 PromptCache）
|
||
export const LLM_PROMPTS = {
|
||
get reflection() { return { system: getSystemPrompt('reflection'), user: getUserPromptTemplate('reflection')! }; },
|
||
get compaction() { return { system: getSystemPrompt('compaction'), user: getUserPromptTemplate('compaction')! }; },
|
||
get extraction() { return { system: getSystemPrompt('extraction'), user: getUserPromptTemplate('extraction')! }; },
|
||
get suggestions() { return { system: getSystemPrompt('suggestions'), user: getUserPromptTemplate('suggestions')! }; },
|
||
};
|
||
|
||
// === Telemetry Integration ===
|
||
|
||
/**
|
||
* 记录一次 LLM 调用结果到遥测收集器。
|
||
* 所有 adapter 的 complete() 返回后应调用此函数。
|
||
*/
|
||
function trackLLMCall(
|
||
adapter: LLMServiceAdapter,
|
||
response: LLMResponse,
|
||
error?: unknown,
|
||
): void {
|
||
try {
|
||
recordLLMUsage(
|
||
response.model || adapter.getProvider(),
|
||
response.tokensUsed?.input ?? 0,
|
||
response.tokensUsed?.output ?? 0,
|
||
{
|
||
latencyMs: response.latencyMs,
|
||
success: !error,
|
||
errorType: error instanceof Error ? error.message.slice(0, 80) : undefined,
|
||
connectionMode: adapter.getProvider() === 'saas' ? 'saas' : 'tauri',
|
||
},
|
||
);
|
||
} catch (e) {
|
||
log.debug('Telemetry recording failed', { error: e });
|
||
}
|
||
}
|
||
|
||
// === Helper Functions ===
|
||
|
||
export async function llmReflect(context: string, adapter?: LLMServiceAdapter): Promise<string> {
|
||
const llm = adapter || getLLMAdapter();
|
||
|
||
const response = await llm.complete([
|
||
{ role: 'system', content: LLM_PROMPTS.reflection.system },
|
||
{ role: 'user', content: typeof LLM_PROMPTS.reflection.user === 'function' ? LLM_PROMPTS.reflection.user(context) : LLM_PROMPTS.reflection.user },
|
||
]);
|
||
|
||
trackLLMCall(llm, response);
|
||
return response.content;
|
||
}
|
||
|
||
export async function llmCompact(messages: string, adapter?: LLMServiceAdapter): Promise<string> {
|
||
const llm = adapter || getLLMAdapter();
|
||
|
||
const response = await llm.complete([
|
||
{ role: 'system', content: LLM_PROMPTS.compaction.system },
|
||
{ role: 'user', content: typeof LLM_PROMPTS.compaction.user === 'function' ? LLM_PROMPTS.compaction.user(messages) : LLM_PROMPTS.compaction.user },
|
||
]);
|
||
|
||
trackLLMCall(llm, response);
|
||
return response.content;
|
||
}
|
||
|
||
export async function llmExtract(
|
||
conversation: string,
|
||
adapter?: LLMServiceAdapter
|
||
): Promise<string> {
|
||
const llm = adapter || getLLMAdapter();
|
||
|
||
const response = await llm.complete([
|
||
{ role: 'system', content: LLM_PROMPTS.extraction.system },
|
||
{ role: 'user', content: typeof LLM_PROMPTS.extraction.user === 'function' ? LLM_PROMPTS.extraction.user(conversation) : LLM_PROMPTS.extraction.user },
|
||
]);
|
||
|
||
trackLLMCall(llm, response);
|
||
return response.content;
|
||
}
|
||
|
||
export async function llmSuggest(
|
||
conversationContext: string,
|
||
adapter?: LLMServiceAdapter,
|
||
): Promise<string> {
|
||
const llm = adapter || getLLMAdapter();
|
||
|
||
const response = await llm.complete([
|
||
{ role: 'system', content: LLM_PROMPTS.suggestions.system },
|
||
{ role: 'user', content: typeof LLM_PROMPTS.suggestions.user === 'function' ? LLM_PROMPTS.suggestions.user(conversationContext) : LLM_PROMPTS.suggestions.user },
|
||
]);
|
||
|
||
trackLLMCall(llm, response);
|
||
return response.content;
|
||
}
|