1.ReAct模式——推理+行動
ReAct（理性 + 行動）是代理 AI 最強的模式 — LLM 行動前思考, 和 收到結果後回饋。與簡單的函數呼叫不同，ReAct 創建思想痕跡來幫助偵錯和改進代理。
┌────────────── ReAct LOOP ──────────────────┐
│ │
│ ┌──────────┐ │
│ │ Thought │ "Tôi cần kiểm tra đơn hàng" │
│ └────┬─────┘ │
│ │ │
│ ┌────▼─────┐ │
│ │ Action │ call get_order(id="12345") │
│ └────┬─────┘ │
│ │ │
│ ┌────▼──────┐ │
│ │Observation│ {status: "shipped", ...} │
│ └────┬──────┘ │
│ │ │
│ ┌────▼─────┐ │
│ │ Thought │ "Đơn đang ship, cần kiểm │
│ │ │ tra tracking number" │
│ └────┬─────┘ │
│ │ │
│ ┌────▼─────┐ │
│ │ Action │ call get_tracking(...) │
│ └────┬─────┘ │
│ │ │
│ ┌────▼──────┐ │
│ │ Answer │ "Đơn #12345 đang ship, │
│ │ │ mã tracking: VN123..." │
│ └──────────┘ │
└─────────────────────────────────────────────┘
class ReActAgent {
private maxIterations = 10;
async execute(
task: string,
tools: ToolDefinition[],
context: AgentContext,
): Promise<ReActResult> {
const trace: ReActStep[] = [];
const systemPrompt = `
You are a reasoning agent. For each step, you MUST follow this format:
Thought: [Your reasoning about what to do next]
Action: [Tool name to call, or "answer" if you have enough info]
Action Input: [JSON arguments for the tool]
After receiving an observation, reason about the result before taking the next action.
If you have enough information to answer, use Action: answer.
Available tools:
${tools.map(t => `- ${t.name}: ${t.description}`).join('\n')}
`;
const messages: LLMMessage[] = [
{ role: 'system', content: systemPrompt },
{ role: 'user', content: task },
];
for (let i = 0; i < this.maxIterations; i++) {
// Get LLM's thought + action
const response = await this.llm.chat({
messages,
model: context.model,
temperature: 0.2,
stop: ['Observation:'], // Stop before observation
});
const parsed = this.parseReActResponse(response.content);
trace.push(parsed);
// If agent wants to answer, we're done
if (parsed.action === 'answer') {
return {
answer: parsed.actionInput as string,
trace,
iterations: i + 1,
};
}
// Execute tool
const toolResult = await this.toolExecutor.execute(
{ name: parsed.action, arguments: parsed.actionInput },
context,
);
// Add observation
messages.push({ role: 'assistant', content: response.content });
messages.push({
role: 'user',
content: `Observation: ${JSON.stringify(toolResult.data ?? toolResult.error)}`,
});
}
return {
answer: 'Không thể hoàn thành task trong số bước cho phép.',
trace,
iterations: this.maxIterations,
};
}
private parseReActResponse(text: string): ReActStep {
const thoughtMatch = text.match(/Thought:\s*(.+?)(?=\nAction:)/s);
const actionMatch = text.match(/Action:\s*(.+?)(?=\nAction Input:)/s);
const inputMatch = text.match(/Action Input:\s*(.+)/s);
return {
thought: thoughtMatch?.[1]?.trim() ?? '',
action: actionMatch?.[1]?.trim() ?? 'answer',
actionInput: inputMatch ? JSON.parse(inputMatch[1].trim()) : text,
};
}
}
2. 計劃-執行-審查週期
interface TaskPlan {
goal: string;
steps: PlanStep[];
dependencies: Map<number, number[]>; // step → depends on steps
}
interface PlanStep {
id: number;
description: string;
tool: string;
expectedOutput: string;
status: 'pending' | 'executing' | 'completed' | 'failed' | 'skipped';
result?: unknown;
}
class PlanExecuteAgent {
async execute(task: string, context: AgentContext): Promise<PlanResult> {
// Phase 1: PLAN
const plan = await this.createPlan(task, context);
// Phase 2: EXECUTE step by step
for (const step of plan.steps) {
// Check dependencies
const deps = plan.dependencies.get(step.id) ?? [];
const depsCompleted = deps.every(
d => plan.steps.find(s => s.id === d)?.status === 'completed',
);
if (!depsCompleted) {
step.status = 'skipped';
continue;
}
step.status = 'executing';
try {
const result = await this.executeStep(step, plan, context);
step.result = result;
step.status = 'completed';
} catch (error) {
step.status = 'failed';
// Phase 3: REPLAN on failure
const replan = await this.replan(plan, step, error, context);
if (replan) {
plan.steps = replan.steps;
continue;
}
}
}
// Phase 4: REVIEW — Self-assess the results
const review = await this.reviewResults(task, plan, context);
return { plan, review, finalAnswer: review.answer };
}
private async createPlan(task: string, context: AgentContext): Promise<TaskPlan> {
const response = await this.llm.chat({
messages: [{
role: 'system',
content: `Create a step-by-step plan to accomplish the task.
Output JSON:
{
"goal": "...",
"steps": [
{"id": 1, "description": "...", "tool": "tool_name", "expectedOutput": "..."},
...
],
"dependencies": {"2": [1], "3": [1, 2]}
}
Available tools: ${context.tools.map(t => t.name).join(', ')}`,
}, {
role: 'user',
content: task,
}],
response_format: { type: 'json_object' },
});
return JSON.parse(response.content);
}
private async reviewResults(
task: string,
plan: TaskPlan,
context: AgentContext,
): Promise<ReviewResult> {
const response = await this.llm.chat({
messages: [{
role: 'system',
content: `Review if the plan execution successfully addressed the user's task.
Assess:
1. Were all necessary steps completed?
2. Is the information sufficient to answer?
3. Are there any inconsistencies?
Provide a final answer to the user.`,
}, {
role: 'user',
content: `Task: ${task}\n\nPlan results:\n${JSON.stringify(plan.steps, null, 2)}`,
}],
});
return { answer: response.content, isComplete: true };
}
}
3.自我反思與錯誤修復
class ReflectionAgent {
async executeWithReflection(
task: string,
context: AgentContext,
maxReflections = 3,
): Promise<ReflectionResult> {
let attempt = 0;
let lastResponse = '';
const reflections: Reflection[] = [];
while (attempt < maxReflections) {
// Generate response
const response = await this.generateResponse(task, context, reflections);
lastResponse = response;
// Self-reflect on quality
const reflection = await this.reflect(task, response, context);
reflections.push(reflection);
if (reflection.isAcceptable) {
return {
finalResponse: response,
reflections,
attempts: attempt + 1,
};
}
// Log improvement areas
console.log(`Reflection ${attempt + 1}: ${reflection.feedback}`);
attempt++;
}
// Return best attempt with disclaimer
return {
finalResponse: lastResponse,
reflections,
attempts: maxReflections,
disclaimer: 'Response may be incomplete — reached max reflection attempts.',
};
}
private async reflect(
task: string,
response: string,
context: AgentContext,
): Promise<Reflection> {
const result = await this.llm.chat({
messages: [{
role: 'system',
content: `You are a quality reviewer. Evaluate the response against the task.
Check for:
1. Accuracy: Does it correctly address the task?
2. Completeness: Are all aspects covered?
3. Hallucination: Any made-up information?
4. Consistency: Does it contradict known facts?
5. Actionability: Can the user act on this?
Output JSON:
{
"isAcceptable": true/false,
"score": 1-10,
"issues": ["issue1", ...],
"feedback": "specific improvement suggestions",
"hallucinations": ["any detected hallucinations"]
}`,
}, {
role: 'user',
content: `Task: ${task}\n\nResponse to evaluate:\n${response}`,
}],
response_format: { type: 'json_object' },
model: 'gpt-4o', // Use strong model for reflection
});
return JSON.parse(result.content);
}
}
4. 幻覺偵測
class HallucinationDetector {
async check(
response: string,
sources: Document[],
toolResults: ToolResult[],
): Promise<HallucinationReport> {
// Strategy 1: Source attribution check
const claims = await this.extractClaims(response);
const unsupported: string[] = [];
for (const claim of claims) {
const isSupported = await this.isClaimSupported(claim, sources, toolResults);
if (!isSupported) unsupported.push(claim);
}
// Strategy 2: Self-consistency check (ask same question N times)
const consistencyScore = await this.checkConsistency(response);
// Strategy 3: Factual verification (for numeric claims)
const numericClaims = this.extractNumericClaims(response);
const numericErrors = await this.verifyNumericClaims(numericClaims, sources);
return {
overallScore: this.calculateScore(unsupported, consistencyScore, numericErrors),
unsupportedClaims: unsupported,
consistencyScore,
numericErrors,
recommendation: unsupported.length > 0 ? 'FLAG_FOR_REVIEW' : 'PASS',
};
}
private async extractClaims(text: string): Promise<string[]> {
const response = await this.llm.chat({
messages: [{
role: 'system',
content: 'Extract all factual claims from the text. Output as JSON array of strings.',
}, {
role: 'user',
content: text,
}],
response_format: { type: 'json_object' },
model: 'gpt-4o-mini',
});
return JSON.parse(response.content).claims;
}
private async isClaimSupported(
claim: string,
sources: Document[],
toolResults: ToolResult[],
): Promise<boolean> {
// Check against RAG sources
const claimEmbedding = await this.embedder.embed(claim);
for (const source of sources) {
const sourceEmbedding = await this.embedder.embed(source.content);
const similarity = this.cosineSimilarity(claimEmbedding, sourceEmbedding);
if (similarity > 0.8) return true;
}
// Check against tool results
const allResults = JSON.stringify(toolResults);
if (allResults.toLowerCase().includes(claim.toLowerCase().slice(0, 50))) {
return true;
}
return false;
}
}
5. 信賴度評分框架
interface ConfidenceScore {
overall: number; // 0.0 - 1.0
sourceReliability: number;
modelCertainty: number;
responseConsistency: number;
action: 'respond' | 'caveat' | 'escalate' | 'refuse';
}
class ConfidenceScorer {
score(response: AgentResponse): ConfidenceScore {
const sourceReliability = this.scoreSourceReliability(response.sources);
const modelCertainty = this.scoreModelCertainty(response.logprobs);
const responseConsistency = response.hallucinationReport.consistencyScore;
const overall = (
sourceReliability * 0.4 +
modelCertainty * 0.3 +
responseConsistency * 0.3
);
return {
overall,
sourceReliability,
modelCertainty,
responseConsistency,
action: this.determineAction(overall),
};
}
private determineAction(score: number): ConfidenceScore['action'] {
if (score >= 0.8) return 'respond'; // High confidence
if (score >= 0.5) return 'caveat'; // Add disclaimer
if (score >= 0.3) return 'escalate'; // Transfer to human
return 'refuse'; // Don't answer
}
private scoreModelCertainty(logprobs: number[] | undefined): number {
if (!logprobs?.length) return 0.5;
const avgLogprob = logprobs.reduce((a, b) => a + b, 0) / logprobs.length;
return Math.exp(avgLogprob); // Convert log probability to probability
}
}
第 10 課總結
- 反應：思想→行動→觀察循環－智能體在行動前思考
- 計劃-執行-審查：制定計劃→逐步執行→審核結果→失敗重新規劃
- 自我反省：代理自我評估品質、檢測問題、改進回應（最多 N 次嘗試）
- 幻覺偵測：提取聲明→檢查來源→標記不受支援的聲明
- 置信度評分：≥0.8 回應，≥0.5 新增警告，≥0.3 升級，<0.3 拒絕
下一篇： 結構化資料查詢－文字到 SQL、知識圖查詢、資料庫代理程式、結果格式化。