Partilhar via

Quickstart: Pesquisa vetorial com Python no Azure DocumentDB

Use pesquisa vetorial no Azure DocumentDB com a biblioteca cliente Python. Armazene e consulte dados vetoriais de forma eficiente.

Este início rápido usa um conjunto de dados de hotel de exemplo em um arquivo JSON com vetores do modelo text-embedding-ada-002. O conjunto de dados inclui nomes de hotéis, locais, descrições e incorporações vetoriais.

Encontre o código de exemplo no GitHub.

Pré-requisitos

  • Uma assinatura do Azure

    • Se você não tiver uma assinatura do Azure, crie uma conta gratuita

  • Um cluster do Azure DocumentDB existente

Criar um projeto Python

  1. Crie um novo diretório para seu projeto e abra-o no Visual Studio Code:

    mkdir vector-search-quickstart
code vector-search-quickstart

  2. No terminal, crie e ative um ambiente virtual:

    Para Windows:

    python -m venv venv
venv\\Scripts\\activate

    Para macOS/Linux:

    python -m venv venv
source venv/bin/activate

  3. Instale os pacotes necessários:

    pip install pymongo azure-identity openai python-dotenv
    • pymongo: Driver MongoDB para Python
    • azure-identity: Biblioteca de Identidade do Azure para autenticação sem senha
    • openai: Biblioteca de cliente OpenAI para criar vetores
    • python-dotenv: Gerenciamento de variáveis de ambiente a partir de arquivos .env

  4. Crie um arquivo .env na raiz do seu projeto para variáveis de ambiente.

    # Azure OpenAI configuration
AZURE_OPENAI_EMBEDDING_ENDPOINT= 
AZURE_OPENAI_EMBEDDING_MODEL=text-embedding-ada-002
AZURE_OPENAI_EMBEDDING_API_VERSION=2024-02-01

# Azure DocumentDB configuration
MONGO_CLUSTER_NAME=

# Data Configuration (defaults should work)
DATA_FILE_WITH_VECTORS=data/HotelsData_with_vectors.json
EMBEDDED_FIELD=text_embedding_ada_002
EMBEDDING_DIMENSIONS=1536
EMBEDDING_SIZE_BATCH=16
LOAD_SIZE_BATCH=100

    Para a autenticação sem senha usada neste artigo, substitua os valores placeholders .env no arquivo por suas próprias informações.

    • AZURE_OPENAI_EMBEDDING_ENDPOINT: URL do ponto final do recurso Azure OpenAI
    • MONGO_CLUSTER_NAME: Nome do recurso do Azure DocumentDB

    Você deve sempre preferir a autenticação sem senha, mas ela exigirá configuração adicional. Para obter mais informações sobre como configurar a identidade gerenciada e a gama completa de suas opções de autenticação, consulte Autenticar aplicativos Python nos serviços do Azure usando o SDK do Azure para Python.

  5. Crie um novo subdiretório fora da raiz chamado data.

  6. Copie o arquivo de dados brutos com vetores para um novo HotelsData_with_vectors.json arquivo no data subdiretório.

  7. A estrutura do projeto deve ter esta aparência:

    vector-search-quickstart
├── .env
├── data
│   └── HotelsData_with_vectors.json
└── venv (or your virtual environment folder)

Continue o projeto criando arquivos de código para pesquisa vetorial. Quando terminar, a estrutura do projeto deve ter esta aparência:

vector-search-quickstart
├── .env
├── data
│   └── HotelsData_with_vectors.json
├── src
│   ├── diskann.py
│   ├── ivf.py
│   └── hnsw.py
│   └── utils.py
└── venv (or your virtual environment folder)

Crie um src diretório para seus arquivos Python. Adicione dois arquivos: diskann.py e utils.py para a implementação do índice DiskANN:

mkdir src    
touch src/diskann.py
touch src/utils.py

Cole o código a seguir no diskann.py arquivo.

import os
from typing import List, Dict, Any
from utils import get_clients, get_clients_passwordless, read_file_return_json, insert_data, print_search_results, drop_vector_indexes
from dotenv import load_dotenv

# Load environment variables
load_dotenv()


def create_diskann_vector_index(collection, vector_field: str, dimensions: int) -> None:

    print(f"Creating DiskANN vector index on field '{vector_field}'...")

    # Drop any existing vector indexes on this field first
    drop_vector_indexes(collection, vector_field)

    # Use the native MongoDB command for Cosmos DB vector indexes
    index_command = {
        "createIndexes": collection.name,
        "indexes": [
            {
                "name": f"diskann_index_{vector_field}",
                "key": {
                    vector_field: "cosmosSearch"  # Cosmos DB vector search index type
                },
                "cosmosSearchOptions": {
                    # DiskANN algorithm configuration
                    "kind": "vector-diskann",

                    # Vector dimensions must match the embedding model
                    "dimensions": dimensions,

                    # Vector similarity metric - cosine is good for text embeddings
                    "similarity": "COS",

                    # Maximum degree: number of edges per node in the graph
                    # Higher values improve accuracy but increase memory usage
                    "maxDegree": 20,

                    # Build parameter: candidates evaluated during index construction
                    # Higher values improve index quality but increase build time
                    "lBuild": 10
                }
            }
        ]
    }

    try:
        # Execute the createIndexes command directly
        result = collection.database.command(index_command)
        print("DiskANN vector index created successfully")

    except Exception as e:
        print(f"Error creating DiskANN vector index: {e}")

        # Check if it's a tier limitation and suggest alternatives
        if "not enabled for this cluster tier" in str(e):
            print("\nDiskANN indexes require a higher cluster tier.")
            print("Try one of these alternatives:")
            print("  • Upgrade your Cosmos DB cluster to a higher tier")
            print("  • Use HNSW instead: python src/hnsw.py")
            print("  • Use IVF instead: python src/ivf.py")
        raise


def perform_diskann_vector_search(collection,
                                 azure_openai_client,
                                 query_text: str,
                                 vector_field: str,
                                 model_name: str,
                                 top_k: int = 5) -> List[Dict[str, Any]]:

    print(f"Performing DiskANN vector search for: '{query_text}'")

    try:
        # Generate embedding for the query text
        embedding_response = azure_openai_client.embeddings.create(
            input=[query_text],
            model=model_name
        )

        query_embedding = embedding_response.data[0].embedding

        # Construct the aggregation pipeline for vector search
        # Cosmos DB for MongoDB vCore uses $search with cosmosSearch
        pipeline = [
            {
                "$search": {
                    # Use cosmosSearch for vector operations in Cosmos DB
                    "cosmosSearch": {
                        # The query vector to search for
                        "vector": query_embedding,

                        # Field containing the document vectors to compare against
                        "path": vector_field,

                        # Number of final results to return
                        "k": top_k
                    }
                }
            },
            {
                # Add similarity score to the results
                "$project": {
                    "document": "$$ROOT",
                    # Add search score from metadata
                    "score": {"$meta": "searchScore"}
                }
            }
        ]

        # Execute the aggregation pipeline
        results = list(collection.aggregate(pipeline))

        return results

    except Exception as e:
        print(f"Error performing DiskANN vector search: {e}")
        raise


def main():

    # Load configuration from environment variables
    config = {
        'cluster_name': os.getenv('MONGO_CLUSTER_NAME', 'vectorSearch'),
        'database_name': 'vectorSearchDB',
        'collection_name': 'vectorSearchCollection',
        'data_file': os.getenv('DATA_FILE_WITH_VECTORS', 'data/HotelsData_with_vectors.json'),
        'vector_field': os.getenv('EMBEDDED_FIELD', 'DescriptionVector'),
        'model_name': os.getenv('AZURE_OPENAI_EMBEDDING_MODEL', 'text-embedding-ada-002'),
        'dimensions': int(os.getenv('EMBEDDING_DIMENSIONS', '1536')),
        'batch_size': int(os.getenv('LOAD_SIZE_BATCH', '100'))
    }

    try:
        # Initialize clients
        print("\nInitializing MongoDB and Azure OpenAI clients...")
        mongo_client, azure_openai_client = get_clients_passwordless()

        # Get database and collection
        database = mongo_client[config['database_name']]
        collection = database[config['collection_name']]

        # Load data with embeddings
        print(f"\nLoading data from {config['data_file']}...")
        data = read_file_return_json(config['data_file'])
        print(f"Loaded {len(data)} documents")

        # Verify embeddings are present
        documents_with_embeddings = [doc for doc in data if config['vector_field'] in doc]
        if not documents_with_embeddings:
            raise ValueError(f"No documents found with embeddings in field '{config['vector_field']}'. "
                           "Please run create_embeddings.py first.")

        # Insert data into collection
        print(f"\nInserting data into collection '{config['collection_name']}'...")

        # Clear existing data to ensure clean state
        collection.delete_many({})
        print("Cleared existing data from collection")

        # Insert the hotel data
        stats = insert_data(
            collection,
            documents_with_embeddings,
            batch_size=config['batch_size']
        )

        if stats['inserted'] == 0:
            raise ValueError("No documents were inserted successfully")

        # Create DiskANN vector index
        create_diskann_vector_index(
            collection,
            config['vector_field'],
            config['dimensions']
        )

        # Wait briefly for index to be ready
        import time
        print("Waiting for index to be ready...")
        time.sleep(2)

        # Perform sample vector search
        query = "quintessential lodging near running trails, eateries, retail"

        results = perform_diskann_vector_search(
            collection,
            azure_openai_client,
            query,
            config['vector_field'],
            config['model_name'],
            top_k=5
        )

        # Display results
        print_search_results(results, max_results=5, show_score=True)


    except Exception as e:
        print(f"\nError during DiskANN demonstration: {e}")
        raise

    finally:
        # Close the MongoDB client
        if 'mongo_client' in locals():
            mongo_client.close()


if __name__ == "__main__":
    main()

Este módulo principal fornece estas características:

  • Inclui funções utilitárias

  • Cria um objeto de configuração para variáveis de ambiente

  • Cria clientes para Azure OpenAI e Azure DocumentDB

  • Conecta-se ao MongoDB, cria um banco de dados e uma coleção, insere dados e cria índices padrão

  • Cria um índice vetorial usando FIV, HNSW ou DiskANN

  • Cria uma incorporação para um texto de consulta de exemplo usando o cliente OpenAI. Você pode alterar a consulta na parte superior do arquivo

  • Executa uma pesquisa vetorial usando a incorporação e imprime os resultados

Criar funções utilitárias

Cole o seguinte código em utils.py:

import json
import os
import time
from typing import Dict, List, Any, Optional, Tuple
from pymongo import MongoClient, InsertOne
from pymongo.collection import Collection
from pymongo.errors import BulkWriteError
from azure.identity import DefaultAzureCredential
from pymongo.auth_oidc import OIDCCallback, OIDCCallbackContext, OIDCCallbackResult
from openai import AzureOpenAI
from dotenv import load_dotenv

# Load environment variables from .env file
load_dotenv()

class AzureIdentityTokenCallback(OIDCCallback):
    def __init__(self, credential):
        self.credential = credential

    def fetch(self, context: OIDCCallbackContext) -> OIDCCallbackResult:
        token = self.credential.get_token(
            "https://ossrdbms-aad.database.windows.net/.default").token
        return OIDCCallbackResult(access_token=token)

def get_clients() -> Tuple[MongoClient, AzureOpenAI]:

    # Get MongoDB connection string - required for Cosmos DB access
    mongo_connection_string = os.getenv("MONGO_CONNECTION_STRING")
    if not mongo_connection_string:
        raise ValueError("MONGO_CONNECTION_STRING environment variable is required")

    # Create MongoDB client with optimized settings for Cosmos DB
    mongo_client = MongoClient(
        mongo_connection_string,
        maxPoolSize=50,  # Allow up to 50 connections for better performance
        minPoolSize=5,   # Keep minimum 5 connections open
        maxIdleTimeMS=30000,  # Close idle connections after 30 seconds
        serverSelectionTimeoutMS=5000,  # 5 second timeout for server selection
        socketTimeoutMS=20000  # 20 second socket timeout
    )

    # Get Azure OpenAI configuration
    azure_openai_endpoint = os.getenv("AZURE_OPENAI_EMBEDDING_ENDPOINT")
    azure_openai_key = os.getenv("AZURE_OPENAI_EMBEDDING_KEY")

    if not azure_openai_endpoint or not azure_openai_key:
        raise ValueError("Azure OpenAI endpoint and key are required")

    # Create Azure OpenAI client for generating embeddings
    azure_openai_client = AzureOpenAI(
        azure_endpoint=azure_openai_endpoint,
        api_key=azure_openai_key,
        api_version=os.getenv("AZURE_OPENAI_EMBEDDING_API_VERSION", "2024-02-01")
    )

    return mongo_client, azure_openai_client


def get_clients_passwordless() -> Tuple[MongoClient, AzureOpenAI]:

    # Get MongoDB cluster name for passwordless authentication
    cluster_name = os.getenv("MONGO_CLUSTER_NAME")
    if not cluster_name:
        raise ValueError("MONGO_CLUSTER_NAME environment variable is required")

    # Create credential object for Azure authentication
    credential = DefaultAzureCredential()

    authProperties = {"OIDC_CALLBACK": AzureIdentityTokenCallback(credential)}

    # Create MongoDB client with Azure AD token callback
    mongo_client = MongoClient(
        f"mongodb+srv://{cluster_name}.global.mongocluster.cosmos.azure.com/",
        connectTimeoutMS=120000,
        tls=True,
        retryWrites=True,
        authMechanism="MONGODB-OIDC",
        authMechanismProperties=authProperties
    )

    # Get Azure OpenAI endpoint
    azure_openai_endpoint = os.getenv("AZURE_OPENAI_EMBEDDING_ENDPOINT")
    if not azure_openai_endpoint:
        raise ValueError("AZURE_OPENAI_EMBEDDING_ENDPOINT environment variable is required")

    # Create Azure OpenAI client with credential-based authentication
    azure_openai_client = AzureOpenAI(
        azure_endpoint=azure_openai_endpoint,
        azure_ad_token_provider=lambda: credential.get_token("https://cognitiveservices.azure.com/.default").token,
        api_version=os.getenv("AZURE_OPENAI_EMBEDDING_API_VERSION", "2024-02-01")
    )

    return mongo_client, azure_openai_client


def azure_identity_token_callback(credential: DefaultAzureCredential) -> str:

    # Cosmos DB for MongoDB requires this specific scope
    token_scope = "https://cosmos.azure.com/.default"

    # Get token from Azure AD
    token = credential.get_token(token_scope)

    return token.token


def read_file_return_json(file_path: str) -> List[Dict[str, Any]]:

    try:
        with open(file_path, 'r', encoding='utf-8') as file:
            return json.load(file)
    except FileNotFoundError:
        print(f"Error: File '{file_path}' not found")
        raise
    except json.JSONDecodeError as e:
        print(f"Error: Invalid JSON in file '{file_path}': {e}")
        raise


def write_file_json(data: List[Dict[str, Any]], file_path: str) -> None:

    try:
        with open(file_path, 'w', encoding='utf-8') as file:
            json.dump(data, file, indent=2, ensure_ascii=False)
        print(f"Data successfully written to '{file_path}'")
    except IOError as e:
        print(f"Error writing to file '{file_path}': {e}")
        raise


def insert_data(collection: Collection, data: List[Dict[str, Any]],
                batch_size: int = 100, index_fields: Optional[List[str]] = None) -> Dict[str, int]:

    total_documents = len(data)
    inserted_count = 0
    failed_count = 0

    print(f"Starting batch insertion of {total_documents} documents...")

    # Create indexes if specified
    if index_fields:
        for field in index_fields:
            try:
                collection.create_index(field)
                print(f"Created index on field: {field}")
            except Exception as e:
                print(f"Warning: Could not create index on {field}: {e}")

    # Process data in batches to manage memory and error recovery
    for i in range(0, total_documents, batch_size):
        batch = data[i:i + batch_size]
        batch_num = (i // batch_size) + 1
        total_batches = (total_documents + batch_size - 1) // batch_size

        try:
            # Prepare bulk insert operations
            operations = [InsertOne(document) for document in batch]

            # Execute bulk insert
            result = collection.bulk_write(operations, ordered=False)
            inserted_count += result.inserted_count

            print(f"Batch {batch_num} completed: {result.inserted_count} documents inserted")

        except BulkWriteError as e:
            # Handle partial failures in bulk operations
            inserted_count += e.details.get('nInserted', 0)
            failed_count += len(batch) - e.details.get('nInserted', 0)

            print(f"Batch {batch_num} had errors: {e.details.get('nInserted', 0)} inserted, "
                  f"{failed_count} failed")

            # Print specific error details for debugging
            for error in e.details.get('writeErrors', []):
                print(f"  Error: {error.get('errmsg', 'Unknown error')}")

        except Exception as e:
            # Handle unexpected errors
            failed_count += len(batch)
            print(f"Batch {batch_num} failed completely: {e}")

        # Small delay between batches to avoid overwhelming the database
        time.sleep(0.1)

    # Return summary statistics
    stats = {
        'total': total_documents,
        'inserted': inserted_count,
        'failed': failed_count
    }

    return stats


def drop_vector_indexes(collection, vector_field: str) -> None:

    try:
        # Get all indexes for the collection
        indexes = list(collection.list_indexes())

        # Find vector indexes on the specified field
        vector_indexes = []
        for index in indexes:
            if 'key' in index and vector_field in index['key']:
                if index['key'][vector_field] == 'cosmosSearch':
                    vector_indexes.append(index['name'])

        # Drop each vector index found
        for index_name in vector_indexes:
            print(f"Dropping existing vector index: {index_name}")
            collection.drop_index(index_name)

        if vector_indexes:
            print(f"Dropped {len(vector_indexes)} existing vector index(es)")
        else:
            print("No existing vector indexes found to drop")

    except Exception as e:
        print(f"Warning: Could not drop existing vector indexes: {e}")
        # Continue anyway - the error might be that no indexes exist


def print_search_resultsx(results: List[Dict[str, Any]],
                        max_results: int = 5,
                        show_score: bool = True) -> None:

    if not results:
        print("No search results found.")
        return

    print(f"\nSearch Results (showing top {min(len(results), max_results)}):")
    print("=" * 80)

    for i, result in enumerate(results[:max_results], 1):

        # Display hotel name and ID
        print(f"HotelName: {result['HotelName']}, Score: {result['score']:.4f}")

def print_search_results(results: List[Dict[str, Any]],
                        max_results: int = 5,
                        show_score: bool = True) -> None:

    if not results:
        print("No search results found.")
        return

    print(f"\nSearch Results (showing top {min(len(results), max_results)}):")
    print("=" * 80)

    for i, result in enumerate(results[:max_results], 1):

        # Check if results are nested under 'document' (when using $$ROOT)
        if 'document' in result:
            doc = result['document']
        else:
            doc = result

        # Display hotel name and ID
        print(f"HotelName: {doc['HotelName']}, Score: {result['score']:.4f}")


    if len(results) > max_results:
        print(f"\n... and {len(results) - max_results} more results")

Este módulo utilitário fornece estes recursos:

  • JsonData: Interface para a estrutura de dados

  • scoreProperty: Localização da pontuação nos resultados da consulta com base no método de pesquisa vetorial

  • getClients: Cria e devolve clientes para Azure OpenAI e Azure DocumentDB

  • getClientsPasswordless: Cria e devolve clientes para Azure OpenAI e Azure DocumentDB usando autenticação sem palavra-passe. Habilite o RBAC em ambos os recursos e entre na CLI do Azure

  • readFileReturnJson: Lê um arquivo JSON e retorna seu conteúdo como uma matriz de JsonData objetos

  • writeFileJson: Grava uma matriz de JsonData objetos em um arquivo JSON

  • insertData: Insere dados em lotes em uma coleção MongoDB e cria índices padrão em campos especificados

  • printSearchResults: Imprime os resultados de uma pesquisa vetorial, incluindo a pontuação e o nome do hotel

Autenticar com a CLI do Azure

Entre na CLI do Azure antes de executar o aplicativo para que ele possa acessar os recursos do Azure com segurança.

az login

Execute o aplicativo

Para executar os scripts Python:

python src/diskann.py

Você vê os cinco principais hotéis que correspondem à consulta de pesquisa vetorial e suas pontuações de semelhança.

Exibir e gerenciar dados no Visual Studio Code

  1. Selecione a extensão DocumentDB no Visual Studio Code para se ligar à sua conta Azure DocumentDB.

  2. Veja os dados e índices na base de dados de Hotéis.

    Captura de ecrã da extensão DocumentDB a mostrar a coleção Azure DocumentDB.

Limpeza de recursos

Elimina o grupo de recursos, a conta Azure DocumentDB e o recurso Azure OpenAI quando não precisares para evitar custos extra.

Conteúdo relacionado

  • Last updated on