DeepEval使用自定义模型评估RAG实例

#导入依赖包import timeimport requestsimport jsonfrom services.ChatService import ChatServicefrom deepeval.models import DeepEvalBaseLLMfrom deepeval.test_case import LLMTestCasefrom deepeval.metrics import (FaithfulnessMetric,                              ContextualPrecisionMetric,                              ContextualRecallMetric,                              ContextualRelevancyMetric)

#自定义模型class QwenModel(DeepEvalBaseLLM):    def __init__(self):        self.api_key = "fastgpt-*******"        self.base_url = "https://jz-fastgpt-stable.djtest.cn/api/v1"        self.model_name = "qwen-max"     def load_model(self):        return self    def generate(self, prompt: str) -> str:        # 调用 Qwen API 的逻辑        headers = {            "Authorization": f"Bearer {self.api_key}",            "Content-Type": "application/json"        }        payload = {            "model": self.model_name,            "messages": [{"role": "user", "content": prompt}],            "temperature": 0        }        response = requests.post(            f"{self.base_url}/chat/completions",            headers=headers,            data=json.dumps(payload)        )        if response.status_code == 200:            return response.json()["choices"][0]["message"]["content"]        else:            raise RuntimeError(f"API 调用失败: {response.status_code}, {response.text}")    async def a_generate(self, prompt: str) -> str:        # 异步实现（与同步类似）        return self.generate(prompt)    def get_model_name(self):        return self.model_name

#评估代码封装class EvalService:    def get_faithfulness(self,ques:str, response):        # 创建评估模型实例        qwen_model = QwenModel()        faithfulness_metric = FaithfulnessMetric(model=qwen_model)        test_case = self.get_test_case(ques,response)        faithfulness_metric.measure(test_case)        faithfulness = dict()        faithfulness["score"] = faithfulness_metric.score        faithfulness["reason"] = faithfulness_metric.reason        print(f"faithfulness:{faithfulness}")        return faithfulness    def get_contextprecision(self, ques: str, response):        # 创建评估模型实例        qwen_model = QwenModel()        contextprecision_metric = ContextualPrecisionMetric(model=qwen_model)        test_case = self.get_test_case(ques,response)        contextprecision_metric.measure(test_case)        contextprecision = dict()        contextprecision["score"] = contextprecision_metric.score        contextprecision["reason"] = contextprecision_metric.reason        print(f"contextprecision:{contextprecision}")        return contextprecision    def get_contextrecall(self, ques: str, response):        # 创建评估模型实例        qwen_model = QwenModel()        contextrecall_metric = ContextualRecallMetric(model=qwen_model)        test_case = self.get_test_case(ques,response)        contextrecall_metric.measure(test_case)        contextrecall = dict()        contextrecall["score"] = contextrecall_metric.score        contextrecall["reason"] = contextrecall_metric.reason        print(f"contextrecall:{contextrecall}")        return contextrecall    def get_contextrelevant(self, ques: str, response):        # 创建评估模型实例        qwen_model = QwenModel()        contextrelevant_metric = ContextualRelevancyMetric(model=qwen_model)        test_case = self.get_test_case(ques,response)        contextrelevant_metric.measure(test_case)        contextrelevant = dict()        contextrelevant["score"] = contextrelevant_metric.score        contextrelevant["reason"] = contextrelevant_metric.reason        print(f"contextrelevant:{contextrelevant}")        return contextrelevant    def get_test_case(self, ques: str, result):        quote_list = result["responseData"][1]["quoteList"]        retrival_context = []        for quote in quote_list:            retrival_context.append(f"{quote['q']}:{quote['a']}")        context = []        historypreview = result["responseData"][2]["historyPreview"]        for history in historypreview:            context.append(history['value'])        answer = result["choices"][0]["message"]["content"]        # 使用自定义模型进行评估        res_case = LLMTestCase(            input=ques,            actual_output=answer,            expected_output=answer,            context=context,            retrieval_context=retrival_context        )        return res_case

if __name__ == "__main__":    url='https://XXXXXX/api/v1/chat/completions'    key='fastgpt-XXXXXX'
    cr=ChatService(url,key)
    #调用ai应用，得到result    result=cr.question_response("XXX怎么收费？")    es = EvalService()
    es.get_faithfulness("XXX怎么收费？", result)    es.get_contextprecision("XXX怎么收费？", result)    es.get_contextrecall("XXX怎么收费？", result)    es.get_contextrelevant("XXX怎么收费？", result)

faithfulness:{'score': 1.0, 'reason': '实际输出与检索上下文完全一致，没有任何矛盾之处，所以得到了满分1.00的忠实度评分。'}contextprecision:{'score': 1.0, 'reason': '得分为1.00，因为相关的节点（即第一个节点）被正确地排在了最前面。'}contextrecall:{'score': 0.5, 'reason': '分数为0.50，因为虽然节点在检索上下文中提到了'}contextrelevant:{'score': 0.16666666666666666, 'reason': "分数为0.17，因为大部分检索内容并未涉及XXX问题，例如……"}