Confident
用于 LLM 单元测试的 DeepEval 包。通过使用 Confident，任何人都可以通过更快的迭代，使用单元测试和集成测试来构建强大的语言模型。我们为从合成数据创建到测试的迭代的每个步骤提供支持。
在本指南中，我们将演示如何测试和衡量 LLM 的性能。我们将展示如何使用我们的回调来衡量性能，以及如何定义您自己的指标并将其记录到我们的仪表板中。
DeepEval 还提供
- 如何生成合成数据
- 如何衡量性能
- 用于监控和审查长期结果的仪表板
安装和设置
%pip install --upgrade --quiet langchain langchain-openai langchain-community deepeval langchain-chroma
获取 API 凭据
要获取 DeepEval API 凭据，请按照以下步骤操作
- 访问 https://app.confident-ai.com
- 点击“组织”
- 复制 API 密钥。
当您登录时，系统还会要求您设置 implementation 名称。实现名称是描述实现类型所必需的。（想想您想如何称呼您的项目。我们建议使其具有描述性。）
!deepeval login
设置 DeepEval
默认情况下，您可以使用 DeepEvalCallbackHandler 来设置您想要跟踪的指标。但是，目前它对指标的支持有限（更多支持即将添加）。目前它支持
from deepeval.metrics.answer_relevancy import AnswerRelevancy
# Here we want to make sure the answer is minimally relevant
answer_relevancy_metric = AnswerRelevancy(minimum_score=0.5)
开始使用
要使用 DeepEvalCallbackHandler，我们需要 implementation_name。
from langchain_community.callbacks.confident_callback import DeepEvalCallbackHandler
deepeval_callback = DeepEvalCallbackHandler(
implementation_name="langchainQuickstart", metrics=[answer_relevancy_metric]
)
API 参考：DeepEvalCallbackHandler
场景 1：馈入 LLM
然后，您可以使用 OpenAI 将其馈入您的 LLM。
from langchain_openai import OpenAI
llm = OpenAI(
temperature=0,
callbacks=[deepeval_callback],
verbose=True,
openai_api_key="<YOUR_API_KEY>",
)
output = llm.generate(
[
"What is the best evaluation tool out there? (no bias at all)",
]
)
API 参考：OpenAI
LLMResult(generations=[[Generation(text='\n\nQ: What did the fish say when he hit the wall? \nA: Dam.', generation_info={'finish_reason': 'stop', 'logprobs': None})], [Generation(text='\n\nThe Moon \n\nThe moon is high in the midnight sky,\nSparkling like a star above.\nThe night so peaceful, so serene,\nFilling up the air with love.\n\nEver changing and renewing,\nA never-ending light of grace.\nThe moon remains a constant view,\nA reminder of life’s gentle pace.\n\nThrough time and space it guides us on,\nA never-fading beacon of hope.\nThe moon shines down on us all,\nAs it continues to rise and elope.', generation_info={'finish_reason': 'stop', 'logprobs': None})], [Generation(text='\n\nQ. What did one magnet say to the other magnet?\nA. "I find you very attractive!"', generation_info={'finish_reason': 'stop', 'logprobs': None})], [Generation(text="\n\nThe world is charged with the grandeur of God.\nIt will flame out, like shining from shook foil;\nIt gathers to a greatness, like the ooze of oil\nCrushed. Why do men then now not reck his rod?\n\nGenerations have trod, have trod, have trod;\nAnd all is seared with trade; bleared, smeared with toil;\nAnd wears man's smudge and shares man's smell: the soil\nIs bare now, nor can foot feel, being shod.\n\nAnd for all this, nature is never spent;\nThere lives the dearest freshness deep down things;\nAnd though the last lights off the black West went\nOh, morning, at the brown brink eastward, springs —\n\nBecause the Holy Ghost over the bent\nWorld broods with warm breast and with ah! bright wings.\n\n~Gerard Manley Hopkins", generation_info={'finish_reason': 'stop', 'logprobs': None})], [Generation(text='\n\nQ: What did one ocean say to the other ocean?\nA: Nothing, they just waved.', generation_info={'finish_reason': 'stop', 'logprobs': None})], [Generation(text="\n\nA poem for you\n\nOn a field of green\n\nThe sky so blue\n\nA gentle breeze, the sun above\n\nA beautiful world, for us to love\n\nLife is a journey, full of surprise\n\nFull of joy and full of surprise\n\nBe brave and take small steps\n\nThe future will be revealed with depth\n\nIn the morning, when dawn arrives\n\nA fresh start, no reason to hide\n\nSomewhere down the road, there's a heart that beats\n\nBelieve in yourself, you'll always succeed.", generation_info={'finish_reason': 'stop', 'logprobs': None})]], llm_output={'token_usage': {'completion_tokens': 504, 'total_tokens': 528, 'prompt_tokens': 24}, 'model_name': 'text-davinci-003'})
然后，您可以通过调用 is_successful() 方法来检查指标是否成功。
answer_relevancy_metric.is_successful()
# returns True/False
运行后，您应该能够在下面看到我们的仪表板。
场景 2：在没有回调的链中跟踪 LLM
要在没有回调的链中跟踪 LLM，您可以在最后插入它。
我们可以从定义一个简单的链开始，如下所示。
import requests
from langchain.chains import RetrievalQA
from langchain_chroma import Chroma
from langchain_community.document_loaders import TextLoader
from langchain_openai import OpenAI, OpenAIEmbeddings
from langchain_text_splitters import CharacterTextSplitter
text_file_url = "https://raw.githubusercontent.com/hwchase17/chat-your-data/master/state_of_the_union.txt"
openai_api_key = "sk-XXX"
with open("state_of_the_union.txt", "w") as f:
response = requests.get(text_file_url)
f.write(response.text)
loader = TextLoader("state_of_the_union.txt")
documents = loader.load()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_documents(documents)
embeddings = OpenAIEmbeddings(openai_api_key=openai_api_key)
docsearch = Chroma.from_documents(texts, embeddings)
qa = RetrievalQA.from_chain_type(
llm=OpenAI(openai_api_key=openai_api_key),
chain_type="stuff",
retriever=docsearch.as_retriever(),
)
# Providing a new question-answering pipeline
query = "Who is the president?"
result = qa.run(query)
定义链后，您可以手动检查答案相似度。
answer_relevancy_metric.measure(result, query)
answer_relevancy_metric.is_successful()
下一步是什么？
您可以在这里创建您自己的自定义指标。
DeepEval 还提供其他功能，例如能够自动创建单元测试，幻觉测试。
如果您有兴趣，请在此处查看我们的 Github 存储库 https://github.com/confident-ai/deepeval。我们欢迎任何 PR 和关于如何提高 LLM 性能的讨论。