AI应用测试工程2026：如何系统化测试你的LLM应用

"我们怎么测试AI应用？"是2026年AI工程师最常被问到的问题之一。传统软件测试方法在这里只够用一半，另一半需要全新的思路。本文给你一套完整的AI应用测试框架。

一、AI应用测试的特殊挑战传统软件测试的假设：相同输入 → 相同输出AI应用的现实：相同输入 → 概率性输出，且输出质量难以用二进制判断。这带来了几个独特挑战：1.不确定性：temperature > 0时，每次运行结果不同2.质量模糊：“好"的答案没有明确边界3.评估成本高：人工评估准确但昂贵，自动评估快速但可能不准4.回归困难：模型升级可能让以前"好"的输出变"差"5.边界情况多：攻击性输入、超长文本、跨语言输入等## 二、AI测试的四个层次第4层：E2E系统测试 ↑ 完整用户场景的端到端验证第3层：集成测试 ↑ RAG检索质量 + 生成质量的联合评估第2层：组件测试 ↑ Prompt模板、检索器、后处理的单独测试第1层：单元测试 ↑ 工具函数、解析逻辑、过滤规则的确定性测试## 三、第一层：确定性组件的单元测试这部分和传统软件测试完全相同：pythonimport pytestfrom unittest.mock import MagicMock, patch# 测试JSON解析函数def test_json_fixer_handles_trailing_comma(): from app.utils import JSONOutputFixer bad_json = '{"key": "value",}' result = JSONOutputFixer.extract_json(bad_json) assert result == {"key": "value"}def test_json_fixer_returns_none_for_invalid(): from app.utils import JSONOutputFixer result = JSONOutputFixer.extract_json("这不是JSON") assert result is None# 测试文档分块逻辑def test_text_splitter_respects_chunk_size(): from langchain_text_splitters import RecursiveCharacterTextSplitter splitter = RecursiveCharacterTextSplitter(chunk_size=100, chunk_overlap=20) text = "A" * 250 chunks = splitter.split_text(text) for chunk in chunks: assert len(chunk) <= 120 # 允许一定超出# 测试检索元数据过滤def test_metadata_filter_construction(): from app.retriever import build_filter filter_obj = build_filter(category="ai", min_date="2026-01-01") assert filter_obj["category"] == "ai" assert "date" in filter_obj# 测试工具参数验证def test_tool_rejects_invalid_params(): from app.tools import SearchTool tool = SearchTool() with pytest.raises(ValueError, match="query不能为空"): tool.run(query="", max_results=5)## 四、第二层：Prompt模板测试pythonfrom dataclasses import dataclassfrom typing import Callableimport pytest@dataclassclass PromptTestCase: name: str inputs: dict expected_behavior: str # 文字描述期望行为 assert_fn: Callable # 验证函数class PromptTester: """Prompt模板测试框架""" def __init__(self, prompt_template, llm_client, use_cache=True): self.template = prompt_template self.llm = llm_client self.cache = {} if use_cache else None def run_test_case(self, test_case: PromptTestCase) -> dict: """运行单个测试用例""" # 渲染prompt prompt = self.template.render(**test_case.inputs) # 获取LLM响应（带缓存） cache_key = hash(prompt) if self.cache is not None and cache_key in self.cache: response = self.cache[cache_key] else: response = self.llm.complete(prompt) if self.cache is not None: self.cache[cache_key] = response # 运行断言 try: test_case.assert_fn(response) return {"name": test_case.name, "status": "pass", "response": response} except AssertionError as e: return {"name": test_case.name, "status": "fail", "error": str(e), "response": response} def run_all(self, test_cases: list[PromptTestCase]) -> dict: """运行所有测试用例""" results = [self.run_test_case(tc) for tc in test_cases] passed = sum(1 for r in results if r["status"] == "pass") return { "total": len(results), "passed": passed, "failed": len(results) - passed, "pass_rate": passed / len(results) if results else 0, "details": results }# 实际测试用例code_review_test_cases = [ PromptTestCase( name="SQL注入检测", inputs={ "language": "python", "code": 'query = f"SELECT * FROM users WHERE id = {user_id}"' }, expected_behavior="应该识别出SQL注入风险", assert_fn=lambda r: ( "sql" in r.lower() or "injection" in r.lower() or "注入" in r, "未检测到SQL注入问题" ) ), PromptTestCase( name="安全代码不误报", inputs={ "language": "python", "code": 'query = "SELECT * FROM users WHERE id = %s"\ncursor.execute(query, (user_id,))' }, expected_behavior="参数化查询不应该报SQL注入", assert_fn=lambda r: ( "critical" not in r.lower() or "sql" not in r.lower(), "对安全代码产生了误报" ) ), PromptTestCase( name="JSON格式输出", inputs={ "language": "python", "code": "x = 1\ny = x + 1" }, expected_behavior="输出必须是有效的JSON", assert_fn=lambda r: __import__('json').loads(r) )]## 五、第三层：RAG检索质量评估pythonimport jsonfrom typing import NamedTupleimport numpy as npclass RAGEvaluationMetrics(NamedTuple): """RAG评估指标""" context_recall: float # 上下文召回率：黄金答案是否在检索结果中 context_precision: float # 上下文精确率：检索结果有多大比例是相关的 faithfulness: float # 忠实度：生成答案是否基于检索内容 answer_relevancy: float # 答案相关性：答案是否回答了问题class RAGEvaluator: """RAG系统质量评估器""" def __init__(self, llm_judge, embedder): self.llm = llm_judge self.embedder = embedder def evaluate_context_recall(self, ground_truth: str, retrieved_contexts: list[str]) -> float: """评估召回率：黄金答案中的信息是否在检索结果中""" if not retrieved_contexts: return 0.0 # 将所有检索内容合并 combined_context = "\n\n".join(retrieved_contexts) prompt = f"""评估以下检索内容是否包含回答问题所需的信息。 正确答案：{ground_truth}检索到的内容：{combined_context}评分标准：- 1.0：检索内容完全包含了正确答案所需的所有信息- 0.7：检索内容包含了大部分信息- 0.4：检索内容只包含少量相关信息- 0.0：检索内容不包含任何相关信息只输出0-1之间的数字：""" response = self.llm.complete(prompt) try: return float(response.strip()) except: return 0.5 def evaluate_faithfulness(self, answer: str, contexts: list[str]) -> float: """评估忠实度：答案是否只基于提供的上下文""" context_text = "\n\n".join(contexts) prompt = f"""判断以下答案是否完全基于提供的上下文，没有引入外部知识或捏造信息。上下文：{context_text}答案：{answer}评分：- 1.0：答案完全基于上下文，无任何编造- 0.7：大部分基于上下文，有少量推断- 0.4：部分基于上下文，但有明显的外部知识引入- 0.0：答案与上下文无关或完全捏造只输出0-1之间的数字：""" response = self.llm.complete(prompt) try: return float(response.strip()) except: return 0.5 def evaluate_answer_relevancy(self, question: str, answer: str) -> float: """评估答案相关性：答案是否回答了问题""" # 使用embedding计算问题和答案的相似度 q_emb = self.embedder.encode(question) a_emb = self.embedder.encode(answer) # 余弦相似度 similarity = np.dot(q_emb, a_emb) / (np.linalg.norm(q_emb) * np.linalg.norm(a_emb)) return float(similarity) def run_evaluation_suite(self, test_cases: list[dict]) -> dict: """运行完整评估套件""" all_metrics = [] for case in test_cases: question = case["question"] ground_truth = case["ground_truth"] # 运行RAG管道 retrieved = self._retrieve(question) answer = self._generate(question, retrieved) # 计算指标 metrics = RAGEvaluationMetrics( context_recall=self.evaluate_context_recall(ground_truth, retrieved), context_precision=self._evaluate_precision(question, retrieved), faithfulness=self.evaluate_faithfulness(answer, retrieved), answer_relevancy=self.evaluate_answer_relevancy(question, answer) ) all_metrics.append(metrics) # 汇总 return { "context_recall": np.mean([m.context_recall for m in all_metrics]), "context_precision": np.mean([m.context_precision for m in all_metrics]), "faithfulness": np.mean([m.faithfulness for m in all_metrics]), "answer_relevancy": np.mean([m.answer_relevancy for m in all_metrics]), "ragas_score": np.mean([ (m.context_recall + m.context_precision + m.faithfulness + m.answer_relevancy) / 4 for m in all_metrics ]) }## 六、LLM作为评判者（LLM-as-Judge）pythonclass LLMJudge: """使用强模型评判弱模型的输出""" COMPARISON_PROMPT = """你是一个公正的AI应用质量评审专家。请比较以下两个AI回答的质量，判断哪个更好。问题：{question}回答A：{answer_a}回答B：{answer_b}评估维度：1. 准确性（答案是否正确）2. 完整性（是否回答了问题的所有方面）3. 清晰度（是否易于理解）4. 简洁性（是否避免冗余）请输出JSON：{{ "winner": "A" 或 "B" 或 "tie", "confidence": 0-1之间的数字, "reasoning": "简短解释"}}""" SINGLE_SCORE_PROMPT = """你是一个专业的AI质量评审员。请评估以下AI回答的质量（针对给定问题）。问题：{question}回答：{answer}参考答案（如有）：{reference}从以下维度打分（1-5）：- 准确性：{accuracy_desc}- 相关性：{relevance_desc}- 有帮助性：{helpfulness_desc}输出JSON：{{ "accuracy": 1-5, "relevance": 1-5, "helpfulness": 1-5, "overall": 1-5, "feedback": "具体改进建议"}}""" def __init__(self, judge_model="gpt-4o"): self.client = __import__('openai').OpenAI() self.judge_model = judge_model def compare(self, question: str, answer_a: str, answer_b: str) -> dict: """比较两个答案，返回哪个更好""" prompt = self.COMPARISON_PROMPT.format( question=question, answer_a=answer_a, answer_b=answer_b ) response = self.client.chat.completions.create( model=self.judge_model, messages=[{"role": "user", "content": prompt}], temperature=0, response_format={"type": "json_object"} ) return json.loads(response.choices[0].message.content) def score(self, question: str, answer: str, reference: str = "") -> dict: """对单个答案评分""" prompt = self.single_score_prompt(question, answer, reference) response = self.client.chat.completions.create( model=self.judge_model, messages=[{"role": "user", "content": prompt}], temperature=0, response_format={"type": "json_object"} ) return json.loads(response.choices[0].message.content)## 七、回归测试框架pythonclass AIRegressionTestSuite: """AI应用回归测试：检测模型升级或prompt变更是否导致性能下降""" def __init__(self, test_set_path: str, judge: LLMJudge): self.test_set = self._load_test_set(test_set_path) self.judge = judge self.baseline_results = {} def capture_baseline(self, rag_pipeline) -> str: """捕获当前版本的基准结果""" results = {} for case in self.test_set: response = rag_pipeline.query(case["question"]) results[case["id"]] = { "question": case["question"], "response": response, "score": self.judge.score( case["question"], response, case.get("reference_answer", "") ) } baseline_path = f"./baselines/{self._get_timestamp()}.json" with open(baseline_path, 'w', encoding='utf-8') as f: json.dump(results, f, ensure_ascii=False, indent=2) self.baseline_results = results return baseline_path def run_regression(self, new_pipeline, baseline_path: str = None) -> dict: """运行回归测试，比较新旧版本""" if baseline_path: with open(baseline_path, 'r', encoding='utf-8') as f: self.baseline_results = json.load(f) regression_results = [] for case in self.test_set: case_id = case["id"] baseline = self.baseline_results.get(case_id) if not baseline: continue # 获取新版本响应 new_response = new_pipeline.query(case["question"]) # 比较新旧版本 comparison = self.judge.compare( question=case["question"], answer_a=baseline["response"], # 旧版本 answer_b=new_response # 新版本 ) regression_results.append({ "case_id": case_id, "question": case["question"], "old_response": baseline["response"][:200], "new_response": new_response[:200], "comparison": comparison, "regression": comparison["winner"] == "A" # 旧版本更好=回归 }) # 统计 regressions = [r for r in regression_results if r["regression"]] improvements = [r for r in regression_results if r["comparison"]["winner"] == "B"] return { "total_cases": len(regression_results), "regressions": len(regressions), "improvements": len(improvements), "neutral": len(regression_results) - len(regressions) - len(improvements), "regression_rate": len(regressions) / len(regression_results) if regression_results else 0, "details": regression_results, "recommendation": "通过" if len(regressions) / len(regression_results) < 0.1 else "需要人工审查" }## 八、CI/CD集成yaml# .github/workflows/ai-tests.ymlname: AI Application Testson: push: branches: [main, develop] pull_request: branches: [main]jobs: unit-tests: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 - uses: actions/setup-python@v5 with: {python-version: '3.11'} - run: pip install pytest pytest-asyncio - run: pytest tests/unit/ -v prompt-tests: needs: unit-tests runs-on: ubuntu-latest env: OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} steps: - uses: actions/checkout@v4 - run: pip install -r requirements.txt - run: pytest tests/prompt/ -v --max-failures=5 # 允许少量失败（AI的不确定性） rag-evaluation: needs: prompt-tests runs-on: ubuntu-latest if: github.event_name == 'pull_request' steps: - uses: actions/checkout@v4 - run: | python scripts/evaluate_rag.py \ --test-set tests/rag_test_cases.json \ --threshold-recall 0.8 \ --threshold-faithfulness 0.85 - name: 评估结果评论到PR uses: actions/github-script@v7 with: script: | const fs = require('fs') const results = JSON.parse(fs.readFileSync('rag_results.json')) github.rest.issues.createComment({ issue_number: context.issue.number, owner: context.repo.owner, repo: context.repo.repo, body: `## RAG评估结果\n\`\`\`json\n${JSON.stringify(results, null, 2)}\n\`\`\`` })## 九、总结：AI测试的核心原则1.确定性部分充分测试：工具函数、解析逻辑等必须有高覆盖率的单元测试2.接受概率性：对于LLM输出，测试"属性"而不是"精确值”（如：包含关键词、格式正确、没有有害内容）3.黄金测试集是资产：维护一套高质量的人工标注测试集，作为评估基准4.LLM-as-Judge：用强模型评判弱模型，比人工评判更快，比规则匹配更准确5.回归测试是安全网：每次模型升级或Prompt变更，必须跑回归测试6.在CI中集成：自动化测试是可持续AI工程的基础，不要等到上线才发现问题AI应用测试是一个新兴领域，最佳实践还在快速演化。但核心原则不变：越早发现问题，修复成本越低。