UpTrain

UpTrain [GitHub || 网站 || 文档] 是一个开源平台，用于评估和改进 LLM 应用程序。它提供了20多个预配置检查（涵盖语言、代码、嵌入用例）的评分，对失败案例进行根本原因分析，并提供解决这些问题的指导。

UpTrain 回调处理器

本 Notebook 展示了 UpTrain 回调处理器如何无缝集成到您的管道中，从而促进多样化的评估。我们选择了一些我们认为适合评估链的评估项。这些评估项会自动运行，结果将显示在输出中。有关 UpTrain 评估的更多详细信息，请参见此处。

此处突出显示了 LangChain 中的选定检索器，用于演示

1. 基础 RAG:

RAG 在检索上下文和生成响应方面起着至关重要的作用。为确保其性能和响应质量，我们进行以下评估：

• 上下文相关性: 确定从查询中提取的上下文是否与响应相关。
• 事实准确性: 评估 LLM 是否产生幻觉或提供不正确的信息。
• 响应完整性: 检查响应是否包含查询请求的所有信息。

2. 多查询生成:

MultiQueryRetriever 会创建与原始问题含义相似的多个问题变体。考虑到其复杂性，我们除了包含之前的评估项外，还增加了：

• 多查询准确性: 确保生成的多个查询与原始查询含义相同。

3. 上下文压缩与重排序:

重排序涉及根据与查询的相关性重新排列节点并选择前 n 个节点。由于重排序完成后节点数量可能会减少，我们执行以下评估：

• 上下文重排序: 检查重排序后的节点顺序是否比原始顺序更与查询相关。
• 上下文简洁性: 检查减少的节点数量是否仍能提供所有所需信息。

这些评估共同确保了 RAG、MultiQueryRetriever 和链中重排序过程的鲁棒性和有效性。

安装依赖项

%pip install -qU langchain langchain_openai langchain-community uptrain faiss-cpu flashrank

huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
	- Avoid using `tokenizers` before the fork if possible
	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
``````output
WARNING: There was an error checking the latest version of pip.
Note: you may need to restart the kernel to use updated packages.

注意：如果您想使用支持 GPU 的库版本，也可以安装 faiss-gpu 而不是 faiss-cpu。

导入库

from getpass import getpass

from langchain.chains import RetrievalQA
from langchain.retrievers import ContextualCompressionRetriever
from langchain.retrievers.document_compressors import FlashrankRerank
from langchain.retrievers.multi_query import MultiQueryRetriever
from langchain_community.callbacks.uptrain_callback import UpTrainCallbackHandler
from langchain_community.document_loaders import TextLoader
from langchain_community.vectorstores import FAISS
from langchain_core.output_parsers.string import StrOutputParser
from langchain_core.prompts.chat import ChatPromptTemplate
from langchain_core.runnables.passthrough import RunnablePassthrough
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_text_splitters import (
    RecursiveCharacterTextSplitter,
)

API 参考：RetrievalQA | ContextualCompressionRetriever | FlashrankRerank | MultiQueryRetriever | UpTrainCallbackHandler | TextLoader | FAISS | StrOutputParser | ChatPromptTemplate | RunnablePassthrough | ChatOpenAI | OpenAIEmbeddings | RecursiveCharacterTextSplitter

加载文档

loader = TextLoader("../../how_to/state_of_the_union.txt")
documents = loader.load()

将文档分割成块

text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
chunks = text_splitter.split_documents(documents)

创建检索器

embeddings = OpenAIEmbeddings()
db = FAISS.from_documents(chunks, embeddings)
retriever = db.as_retriever()

定义 LLM

llm = ChatOpenAI(temperature=0, model="gpt-4")

设置

UpTrain 为您提供

1. 具有高级钻取和筛选选项的仪表板
2. 对失败案例的洞察和常见问题
3. 生产数据的可观测性和实时监控
4. 通过与您的 CI/CD 管道无缝集成进行回归测试

您可以使用以下选项进行 UpTrain 评估：

1. UpTrain 的开源软件 (OSS):

您可以使用开源评估服务来评估您的模型。在这种情况下，您需要提供一个 OpenAI API 密钥。UpTrain 使用 GPT 模型来评估 LLM 生成的响应。您可以在此处获取。

为了在 UpTrain 仪表板中查看您的评估结果，您需要在终端中运行以下命令进行设置：

git clone https://github.com/uptrain-ai/uptrain
cd uptrain
bash run_uptrain.sh

这将在您的本地机器上启动 UpTrain 仪表板。您可以通过 https://:3000/dashboard 访问它。

参数

• key_type="openai"
• api_key="OPENAI_API_KEY"
• project_name="PROJECT_NAME"

2. UpTrain 托管服务与仪表板:

或者，您可以使用 UpTrain 的托管服务来评估您的模型。您可以在此处创建免费的 UpTrain 账户并获得免费试用额度。如果您需要更多试用额度，请在此处与 UpTrain 的维护者预约通话。

使用托管服务的好处是：

1. 无需在本地机器上设置 UpTrain 仪表板。
2. 无需 LLM 的 API 密钥即可访问多个 LLM。

完成评估后，您可以在 https://dashboard.uptrain.ai/dashboard 上的 UpTrain 仪表板中查看结果。

参数

• key_type="uptrain"
• api_key="UPTRAIN_API_KEY"
• project_name="PROJECT_NAME"

注意： project_name 将是 UpTrain 仪表板中显示评估结果的项目名称。

设置 API 密钥

Notebook 将提示您输入 API 密钥。您可以通过更改下方单元格中的 key_type 参数来选择 OpenAI API 密钥或 UpTrain API 密钥。

KEY_TYPE = "openai"  # or "uptrain"
API_KEY = getpass()

1. 基础 RAG

UpTrain 回调处理器将自动捕获生成的查询、上下文和响应，并对响应运行以下三项评估（评分为 0 到 1）：

• 上下文相关性: 检查从查询中提取的上下文是否与响应相关。
• 事实准确性: 检查响应的事实准确性。
• 响应完整性: 检查响应是否包含查询请求的所有信息。

# Create the RAG prompt
template = """Answer the question based only on the following context, which can include text and tables:
{context}
Question: {question}
"""
rag_prompt_text = ChatPromptTemplate.from_template(template)

# Create the chain
chain = (
    {"context": retriever, "question": RunnablePassthrough()}
    | rag_prompt_text
    | llm
    | StrOutputParser()
)

# Create the uptrain callback handler
uptrain_callback = UpTrainCallbackHandler(key_type=KEY_TYPE, api_key=API_KEY)
config = {"callbacks": [uptrain_callback]}

# Invoke the chain with a query
query = "What did the president say about Ketanji Brown Jackson"
docs = chain.invoke(query, config=config)

2024-04-17 17:03:44.969 | INFO     | uptrain.framework.evalllm:evaluate_on_server:378 - Sending evaluation request for rows 0 to <50 to the Uptrain
2024-04-17 17:04:05.809 | INFO     | uptrain.framework.evalllm:evaluate:367 - Local server not running, start the server to log data and visualize in the dashboard!
``````output

Question: What did the president say about Ketanji Brown Jackson
Response: The president mentioned that he had nominated Ketanji Brown Jackson to serve on the United States Supreme Court 4 days ago. He described her as one of the nation's top legal minds who will continue Justice Breyer’s legacy of excellence. He also mentioned that she is a former top litigator in private practice, a former federal public defender, and comes from a family of public school educators and police officers. He described her as a consensus builder and noted that since her nomination, she has received a broad range of support from various groups, including the Fraternal Order of Police and former judges appointed by both Democrats and Republicans.

Context Relevance Score: 1.0
Factual Accuracy Score: 1.0
Response Completeness Score: 1.0

2. 多查询生成

MultiQueryRetriever 用于解决 RAG 管道可能无法根据查询返回最佳文档集的问题。它会生成多个与原始查询含义相同的查询，然后为每个查询获取文档。

为了评估此检索器，UpTrain 将运行以下评估：

• 多查询准确性: 检查生成的多个查询是否与原始查询含义相同。

# Create the retriever
multi_query_retriever = MultiQueryRetriever.from_llm(retriever=retriever, llm=llm)

# Create the uptrain callback
uptrain_callback = UpTrainCallbackHandler(key_type=KEY_TYPE, api_key=API_KEY)
config = {"callbacks": [uptrain_callback]}

# Create the RAG prompt
template = """Answer the question based only on the following context, which can include text and tables:
{context}
Question: {question}
"""
rag_prompt_text = ChatPromptTemplate.from_template(template)

chain = (
    {"context": multi_query_retriever, "question": RunnablePassthrough()}
    | rag_prompt_text
    | llm
    | StrOutputParser()
)

# Invoke the chain with a query
question = "What did the president say about Ketanji Brown Jackson"
docs = chain.invoke(question, config=config)

2024-04-17 17:04:10.675 | INFO     | uptrain.framework.evalllm:evaluate_on_server:378 - Sending evaluation request for rows 0 to <50 to the Uptrain
2024-04-17 17:04:16.804 | INFO     | uptrain.framework.evalllm:evaluate:367 - Local server not running, start the server to log data and visualize in the dashboard!
``````output

Question: What did the president say about Ketanji Brown Jackson
Multi Queries:
  - How did the president comment on Ketanji Brown Jackson?
  - What were the president's remarks regarding Ketanji Brown Jackson?
  - What statements has the president made about Ketanji Brown Jackson?

Multi Query Accuracy Score: 0.5
``````output
2024-04-17 17:04:22.027 | INFO     | uptrain.framework.evalllm:evaluate_on_server:378 - Sending evaluation request for rows 0 to <50 to the Uptrain
2024-04-17 17:04:44.033 | INFO     | uptrain.framework.evalllm:evaluate:367 - Local server not running, start the server to log data and visualize in the dashboard!
``````output

Question: What did the president say about Ketanji Brown Jackson
Response: The president mentioned that he had nominated Circuit Court of Appeals Judge Ketanji Brown Jackson to serve on the United States Supreme Court 4 days ago. He described her as one of the nation's top legal minds who will continue Justice Breyer’s legacy of excellence. He also mentioned that since her nomination, she has received a broad range of support—from the Fraternal Order of Police to former judges appointed by Democrats and Republicans.

Context Relevance Score: 1.0
Factual Accuracy Score: 1.0
Response Completeness Score: 1.0

3. 上下文压缩与重排序

重排序过程涉及根据与查询的相关性重新排列节点并选择前 n 个节点。由于重排序完成后节点数量可能会减少，我们执行以下评估：

• 上下文重排序: 检查重排序后的节点顺序是否比原始顺序更与查询相关。
• 上下文简洁性: 检查减少的节点数量是否仍能提供所有所需信息。

# Create the retriever
compressor = FlashrankRerank()
compression_retriever = ContextualCompressionRetriever(
    base_compressor=compressor, base_retriever=retriever
)

# Create the chain
chain = RetrievalQA.from_chain_type(llm=llm, retriever=compression_retriever)

# Create the uptrain callback
uptrain_callback = UpTrainCallbackHandler(key_type=KEY_TYPE, api_key=API_KEY)
config = {"callbacks": [uptrain_callback]}

# Invoke the chain with a query
query = "What did the president say about Ketanji Brown Jackson"
result = chain.invoke(query, config=config)

2024-04-17 17:04:46.462 | INFO     | uptrain.framework.evalllm:evaluate_on_server:378 - Sending evaluation request for rows 0 to <50 to the Uptrain
2024-04-17 17:04:53.561 | INFO     | uptrain.framework.evalllm:evaluate:367 - Local server not running, start the server to log data and visualize in the dashboard!
``````output

Question: What did the president say about Ketanji Brown Jackson

Context Conciseness Score: 0.0
Context Reranking Score: 1.0
``````output
2024-04-17 17:04:56.947 | INFO     | uptrain.framework.evalllm:evaluate_on_server:378 - Sending evaluation request for rows 0 to <50 to the Uptrain
2024-04-17 17:05:16.551 | INFO     | uptrain.framework.evalllm:evaluate:367 - Local server not running, start the server to log data and visualize in the dashboard!
``````output

Question: What did the president say about Ketanji Brown Jackson
Response: The President mentioned that he nominated Circuit Court of Appeals Judge Ketanji Brown Jackson to serve on the United States Supreme Court 4 days ago. He described her as one of the nation's top legal minds who will continue Justice Breyer’s legacy of excellence.

Context Relevance Score: 1.0
Factual Accuracy Score: 1.0
Response Completeness Score: 0.5

UpTrain 的仪表板和洞察

这里有一个短视频，展示了仪表板和洞察：