利用 Milvus 和 Unstructured 创建 RAG

Unstructured提供了一个平台和工具，用于为检索增强生成（RAG）和模型微调摄取和处理非结构化文档。它提供无代码 UI 平台和无服务器 API 服务，允许用户在 Unstructured 托管的计算资源上处理数据。

在本教程中，我们将使用 Unstructured 采集 PDF 文档，然后使用 Milvus 构建 RAG 管道。

准备工作

依赖和环境

$ pip install -qU "unstructured-ingest[pdf]" unstructured pymilvus openai

您可以从此处获取UNSTRUCTURED_API_KEY 和UNSTRUCTURED_URL 环境变量。

在本例中，我们将使用 OpenAI 作为 LLM。您应将api 密钥 OPENAI_API_KEY 作为环境变量。

import os


os.environ["UNSTRUCTURED_API_KEY"] = "***********"
os.environ["UNSTRUCTURED_URL"] = "***********"

os.environ["OPENAI_API_KEY"] = "***********"

准备 Milvus 和 OpenAI 客户端

您可以使用 Milvus 客户端创建 Milvus Collections 并向其中插入数据。

from pymilvus import MilvusClient, DataType

# Initialize Milvus client
milvus_client = MilvusClient(uri="./milvus_demo.db")  # TODO

检查 Collections 是否已经存在，如果存在，则将其删除。

collection_name = "my_rag_collection"

if milvus_client.has_collection(collection_name):
    milvus_client.drop_collection(collection_name)

准备一个 OpenAI 客户端来生成嵌入和生成响应。

from openai import OpenAI

openai_client = OpenAI()


def emb_text(text):
    return (
        openai_client.embeddings.create(input=text, model="text-embedding-3-small")
        .data[0]
        .embedding
    )

生成一个测试嵌入，并打印其维度和前几个元素。

test_embedding = emb_text("This is a test")
embedding_dim = len(test_embedding)
print(embedding_dim)
print(test_embedding[:10])

1536
[0.009889289736747742, -0.005578675772994757, 0.00683477520942688, -0.03805781528353691, -0.01824733428657055, -0.04121600463986397, -0.007636285852640867, 0.03225184231996536, 0.018949154764413834, 9.352207416668534e-05]

创建 Milvus Collections

我们将创建一个具有以下 Schema 的 Collection：

• id主键：主键是每个文档的唯一标识符。
• vector主键：文档的 Embeddings。
• text文档的文本内容。
• metadata文档的元数据。

然后，我们在vector 字段上建立AUTOINDEX 索引。然后创建 Collections。

# Create schema
schema = milvus_client.create_schema(auto_id=False, enable_dynamic_field=False)
# Add fields to schema
schema.add_field(field_name="id", datatype=DataType.INT64, is_primary=True)
schema.add_field(field_name="vector", datatype=DataType.FLOAT_VECTOR, dim=embedding_dim)
schema.add_field(field_name="text", datatype=DataType.VARCHAR, max_length=65535)
schema.add_field(field_name="metadata", datatype=DataType.JSON)
index_params = MilvusClient.prepare_index_params()
index_params.add_index(
    field_name="vector",
    metric_type="COSINE",
    index_type="AUTOINDEX",
)
milvus_client.create_collection(
    collection_name=collection_name,
    schema=schema,
    index_params=index_params,
    consistency_level="Strong",
)

milvus_client.load_collection(collection_name=collection_name)

从 Unstructured 中加载数据

Unstructured 提供了灵活而强大的摄取管道，可以处理各种类型的文件，包括 PDF、HTML 等。 我们将使用摄取功能来分割本地目录中的 PDF 文件。然后将数据载入 Milvus。

from unstructured_ingest.v2.pipeline.pipeline import Pipeline
from unstructured_ingest.v2.interfaces import ProcessorConfig
from unstructured_ingest.v2.processes.connectors.local import (
    LocalIndexerConfig,
    LocalDownloaderConfig,
    LocalConnectionConfig,
    LocalUploaderConfig,
)
from unstructured_ingest.v2.processes.partitioner import PartitionerConfig

directory_with_pdfs = "./pdf_files"
directory_with_results = "./pdf_processed_outputs"

Pipeline.from_configs(
    context=ProcessorConfig(),
    indexer_config=LocalIndexerConfig(input_path=directory_with_pdfs),
    downloader_config=LocalDownloaderConfig(),
    source_connection_config=LocalConnectionConfig(),
    partitioner_config=PartitionerConfig(
        partition_by_api=True,
        api_key=os.getenv("UNSTRUCTURED_API_KEY"),
        partition_endpoint=os.getenv("UNSTRUCTURED_API_URL"),
        strategy="hi_res",
        additional_partition_args={
            "split_pdf_page": True,
            "split_pdf_concurrency_level": 15,
        },
    ),
    uploader_config=LocalUploaderConfig(output_dir=directory_with_results),
).run()


from unstructured.staging.base import elements_from_json


def load_processed_files(directory_path):
    elements = []
    for filename in os.listdir(directory_path):
        if filename.endswith(".json"):
            file_path = os.path.join(directory_path, filename)
            try:
                elements.extend(elements_from_json(filename=file_path))
            except IOError:
                print(f"Error: Could not read file {filename}.")

    return elements


elements = load_processed_files(directory_with_results)

将数据插入 Milvus。

data = []
for i, element in enumerate(elements):
    data.append(
        {
            "id": i,
            "vector": emb_text(element.text),
            "text": element.text,
            "metadata": element.metadata.to_dict(),
        }
    )
milvus_client.insert(collection_name=collection_name, data=data)

检索和生成响应

定义一个函数，从 Milvus 检索相关文档。

def retrieve_documents(question, top_k=3):
    search_res = milvus_client.search(
        collection_name=collection_name,
        data=[emb_text(question)],
        limit=top_k,
        # search_params={"metric_type": "IP", "params": {}},
        output_fields=["text"],
    )
    return [(res["entity"]["text"], res["distance"]) for res in search_res[0]]

定义一个函数，在 RAG 管道中使用检索到的文档生成响应。

def generate_rag_response(question):
    retrieved_docs = retrieve_documents(question)
    context = "\n".join([f"Text: {doc[0]}\n" for doc in retrieved_docs])
    system_prompt = (
        "You are an AI assistant. Provide answers based on the given context."
    )
    user_prompt = f"""
    Use the following pieces of information to answer the question. If the information is not in the context, say you don't know.
    
    Context:
    {context}
    
    Question: {question}
    """
    response = openai_client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt},
        ],
    )
    return response.choices[0].message.content

让我们用一个示例问题来测试 RAG 管道。

question = "What is the Advanced Search Algorithms in Milvus?"
answer = generate_rag_response(question)
print(f"Question: {question}")
print(f"Answer: {answer}")

INFO: HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
INFO: HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"


Question: What is the Advanced Search Algorithms in Milvus?
Answer: The Advanced Search Algorithms in Milvus refer to a wide range of in-memory and on-disk indexing/search algorithms it supports, including IVF, HNSW, DiskANN, and more. These algorithms have been deeply optimized, and Milvus delivers 30%-70% better performance compared to popular implementations like FAISS and HNSWLib.