可观察性

可观察性有助于跟踪系统的内部工作原理，而无需更改系统。这使得调试和评估整体性能变得更加容易。

txtai 集成了 MLflow 及其跟踪模块，以便深入了解 txtai 中的每个组件。

示例

以下展示了如何在 txtai 进程中引入可观察性的几个示例。

初始化

请先运行以下章节以初始化跟踪。

# Install MLflow plugin for txtai
pip install mlflow-txtai

# Start a local MLflow service
mlflow server --host 127.0.0.1 --port 8000

import mlflow

mlflow.set_tracking_uri(uri="http://localhost:8000")
mlflow.set_experiment("txtai")

# Enable txtai automatic tracing
mlflow.txtai.autolog()

Textractor

第一个示例跟踪 Textractor 管道。

from txtai.pipeline import Textractor

with mlflow.start_run():
    textractor = Textractor()
    textractor("https://github.com/neuml/txtai")

嵌入

接下来，我们将跟踪一个嵌入查询。

from txtai import Embeddings

with mlflow.start_run():
    wiki = Embeddings()
    wiki.load(provider="huggingface-hub", container="neuml/txtai-wikipedia-slim")

    embeddings = Embeddings(content=True, graph=True)
    embeddings.index(wiki.search("SELECT id, text FROM txtai LIMIT 25"))

    embeddings.search("MATCH (A)-[]->(B) RETURN A")

检索增强生成 (RAG)

下一个示例跟踪一个RAG 管道。

from txtai import Embeddings, RAG

with mlflow.start_run():
    wiki = Embeddings()
    wiki.load(provider="huggingface-hub", container="neuml/txtai-wikipedia-slim")

    # Define prompt template
    template = """
    Answer the following question using only the context below. Only include information
    specifically discussed.

    question: {question}
    context: {context} """

    # Create RAG pipeline
    rag = RAG(
        wiki,
        "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
        system="You are a friendly assistant. You answer questions from users.",
        template=template,
        context=10
    )

    rag("Tell me about the Roman Empire", maxlength=2048)

工作流

这个示例运行一个工作流。此工作流运行一个嵌入查询，然后将每个结果翻译成法语。

from txtai import Embeddings, Workflow
from txtai.pipeline import Translation
from txtai.workflow import Task

with mlflow.start_run():
    wiki = Embeddings()
    wiki.load(provider="huggingface-hub", container="neuml/txtai-wikipedia-slim")

    # Translation instance
    translate = Translation()

    workflow = Workflow([
        Task(lambda x: [y[0]["text"] for y in wiki.batchsearch(x, 1)]),
        Task(lambda x: translate(x, "fr"))
    ])

    print(list(workflow(["Roman Empire", "Greek Empire", "Industrial Revolution"])))

代理

最后一个示例运行一个旨在研究天文学问题的 txtai 代理。

from txtai import Agent, Embeddings

def search(query):
    """
    Searches a database of astronomy data.

    Make sure to call this tool only with a string input, never use JSON.    

    Args:
        query: concepts to search for using similarity search

    Returns:
        list of search results with for each match
    """

    return embeddings.search(
        "SELECT id, text, distance FROM txtai WHERE similar(:query)",
        10, parameters={"query": query}
    )

embeddings = Embeddings()
embeddings.load(provider="huggingface-hub", container="neuml/txtai-astronomy")

agent = Agent(
    tools=[search],
    llm="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
    max_iterations=10,
)

researcher = """
{command}

Do the following.
 - Search for results related to the topic.
 - Analyze the results
 - Continue querying until conclusive answers are found
 - Write a Markdown report
"""

with mlflow.start_run():
    agent(researcher.format(command="""
    Write a detailed list with explanations of 10 candidate stars that could potentially be habitable to life.
    """), maxlength=16000)

阅读更多

查看 mlflow-txtai 项目以查看更多示例。