ragas测试

ragas可支持使用不同的 embedding 和 llm 进行计算，默认使用的是 OpenAIEmbeddings(model="text-embedding-ada-002")、ChatOpenAI(model="gpt-3.5-turbo")。

目前国内支持的大模型有：百度千帆、通义千问、百川、Yuan2、智谱ai

下面以三种渠道的embedding 和 llm为例进行ragas测试。不同渠道的api key获取见 API KEY获取和测试

使用ragas+langchain进行rag评估

openai

安装三方库

pip install ragas
pip install datasets 
pip install langchain_openai 

测试代码

import os
from ragas.metrics import answer_relevancy,faithfulness, context_precision,context_relevancy,context_recall
from datasets import Dataset
from ragas import evaluate, RunConfig
from langchain_openai import ChatOpenAI
from langchain_openai import OpenAIEmbeddings
from ragas.llms.base import LangchainLLMWrapper
# 设置环境变量
os.environ["OPENAI_API_BASE"] = "https://api.f2gpt.com"
os.environ["OPENAI_API_KEY"] = "******"

op=ChatOpenAI(model="gpt-4o",temperature=0.5, max_retries=2)
llm = LangchainLLMWrapper(langchain_llm=op)
embeddings=OpenAIEmbeddings()

data_samples = {
    'question': ['四川面积是多少'],
    'answer': ['四川省总面积48万平方千米'],
    'contexts' : [
        ['四川省总面积48.6万平方千米，辖21个[地级行政区]，其中18个[地级市]、3个[自治州]。共55个[市辖区]、19个[县级市]，105个[县]，4个[自治县]，合计183个县级区划。街道459个、镇2016个、乡626个，合计3101个乡级区划。']
      ],
    'ground_truth': ['四川省总面积48.6万平方千米']
}

run_config = RunConfig(
    max_retries=5,
    max_wait=120,
    thread_timeout=500,
    log_tenacity=True
)
dataset = Dataset.from_dict(data_samples)
print("开始计算。。。。。")

# 计算结果 'faithfulness': 0.0000, 'answer_relevancy': nan,, 暂时未找到原因
# 不传llm 和 embeddings，会默认使用 OpenAIEmbeddings(model="text-embedding-ada-002")  ChatOpenAI(model="gpt-3.5-turbo")
score = evaluate(dataset,metrics=[faithfulness,answer_relevancy,context_precision,context_relevancy,context_recall],
                 llm=llm,embeddings=embeddings,
                 raise_exceptions=False,run_config=run_config)

print("计算完成。。。。。。")
print(score)

 

百度千帆

安装三方库

pip install ragas
pip install datasets 
pip install langchain_community
pip install qianfan

测试代码

import os
from ragas.metrics import answer_relevancy,faithfulness, context_precision
from datasets import Dataset
from ragas import evaluate
from langchain_community.llms.baidu_qianfan_endpoint import QianfanLLMEndpoint
from langchain_community.embeddings.baidu_qianfan_endpoint import QianfanEmbeddingsEndpoint
# 设置环境变量
os.environ["QIANFAN_AK"] = "******"
os.environ["QIANFAN_SK"] = "******"

llm = QianfanLLMEndpoint(model="ERNIE-Speed-128K",temperature=0.2,timeout=30)
embeddings=QianfanEmbeddingsEndpoint()

data_samples = {
    'question': ['When was the first super bowl?', 'Who won the most super bowls?'],
    'answer': ['The first superbowl was held on Jan 15, 1967', 'The most super bowls have been won by The New England Patriots'],
    'contexts' : [['The First AFL–NFL World Championship Game was an American football game played on January 15, 1967, at the Los Angeles Memorial Coliseum in Los Angeles,'],
    ['The Green Bay Packers...Green Bay, Wisconsin.','The Packers compete...Football Conference']],
    'ground_truth': ['The first superbowl was held on January 15, 1967', 'The New England Patriots have won the Super Bowl a record six times']
}

dataset = Dataset.from_dict(data_samples)
print("开始计算。。。。。")
# SecretStr str type=type_error.str langchain下一版本修复 todo https://github.com/langchain-ai/langchain/issues/24126
score = evaluate(dataset,metrics=[faithfulness,answer_relevancy,context_precision],llm=llm,embeddings=embeddings,raise_exceptions=False)
print("计算完成。。。。。。")
print(score)

 

通义千问

安装三方库

pip install ragas
pip install datasets 
pip install langchain_community
pip install dashscope

测试代码

import os
from ragas.metrics import faithfulness,answer_relevancy,context_relevancy,context_recall,context_precision,answer_similarity,answer_correctness
from datasets import Dataset
from ragas import evaluate, RunConfig
from langchain_community.llms.tongyi import Tongyi
from langchain_community.embeddings.dashscope import DashScopeEmbeddings

os.environ["DASHSCOPE_API_KEY"] = "******"
llm = Tongyi(model_name="qwen-turbo")
embeddings=DashScopeEmbeddings()

data_samples = {
    'question': ['四川面积是多少','四川有多少人'],
    'answer': ['四川省总面积48.6万平方千米','四川有9071.4万人'],
    'contexts' : [
        ['四川省总面积48.6万平方千米，辖21个[地级行政区]，其中18个[地级市]、3个[自治州]。共55个[市辖区]、19个[县级市]，105个[县]，4个[自治县]，合计183个县级区划。街道459个、镇2016个、乡626个，合计3101个乡级区划。']
      ,['截止2023年末，四川省户籍人口9071.4万人']
    ],
    'ground_truth': ['四川省总面积48.6万平方千米','四川省户籍人口9071.4万人']
}

dataset = Dataset.from_dict(data_samples)
run_config = RunConfig(
    max_retries=5,
    max_wait=120,
    thread_timeout=500,
    log_tenacity=True
)
print("开始计算。。。。。")
score = evaluate(dataset,metrics=[faithfulness,answer_relevancy,context_relevancy,context_recall,context_precision,answer_similarity,answer_correctness],
                 llm=llm,embeddings=embeddings,
                 raise_exceptions=False,run_config=run_config)
print("计算完成。。。。。。")
print(score)

# 评估结果
{'faithfulness': 1.0000, 'answer_relevancy': 0.6457, 'context_relevancy': 0.6667, 'context_recall': 1.0000, 'context_precision': 1.0000, 'answer_similarity': 0.9262, 'answer_correctness': 0.9816}

踩坑记录

1、搭建环境需要c++编译能力(安装 Microsoft Visual Studio，Microsoft Visual C++ ,版本大于2014)，python>3.9

2、网上很多都是安装langchain，但是运行代码都提示 Deprecated，直接安装新版的 langchain_community

3、网上都是建议安装ragas=0.0.22版本，原因都是 langchain 还没有对 ragas 0.1 及以后版本作兼容，发现不需要安装特定版本的ragas，直接安装最新版就行

4、在使用openai渠道测试时，经常出现 faithfulness=0、faithfulness=nan、answer_relevancy=nan的情况，经过排查后仍未找出原因。疑似官方bug。

5、在使用 百度千帆 渠道 测试时，报错 "type=type_error.str，SecretStr！= str"，经过排查后是langchain官方bug，等待修复后发布新版才能使用 https://github.com/langchain-ai/langchain/issues/24126，

6、需要测试多个问题时，data_samples的格式如下：

data_samples = {
    'question': ['question 1', 'question 2', 'question 3'],
    'answer': ['answer to the question 1', 'answer to the question 2', 'answer to the question 3'],
    'contexts': [['context 1'], ['context 2'], ['context 3']],
    'ground_truths': ['ground_truth to the question 1', 'ground_truth to the question 2', 'ground_truth to the question 3']
}