Azure AI Search SDK pour TypeScript

Créez des applications de recherche avec des capacités de recherche vectorielle, hybride et sémantique.

Installation

npm install @azure/search-documents @azure/identity

Variables d'environnement

AZURE_SEARCH_ENDPOINT=https://<service-name>.search.windows.net
AZURE_SEARCH_INDEX_NAME=my-index
AZURE_SEARCH_ADMIN_KEY=<admin-key>  # Optionnel si vous utilisez Entra ID
AZURE_TOKEN_CREDENTIALS=prod # Requis uniquement si DefaultAzureCredential est utilisé en production

Authentification

import { SearchClient, SearchIndexClient } from "@azure/search-documents";
import { DefaultAzureCredential, ManagedIdentityCredential } from "@azure/identity";

const endpoint = process.env.AZURE_SEARCH_ENDPOINT!;
const indexName = process.env.AZURE_SEARCH_INDEX_NAME!;
// Dev local : DefaultAzureCredential. Production : définir AZURE_TOKEN_CREDENTIALS=prod ou AZURE_TOKEN_CREDENTIALS=<specific_credential>
const credential = new DefaultAzureCredential({requiredEnvVars: ["AZURE_TOKEN_CREDENTIALS"]});
// Ou utiliser une credential spécifique directement en production :
// Voir https://learn.microsoft.com/javascript/api/overview/azure/identity-readme?view=azure-node-latest#credential-classes
// const credential = new ManagedIdentityCredential();

// Pour la recherche
const searchClient = new SearchClient(endpoint, indexName, credential);

// Pour la gestion d'index
const indexClient = new SearchIndexClient(endpoint, credential);

Flux de travail principal

Créer un index avec champ vectoriel

import { SearchIndex, SearchField, VectorSearch } from "@azure/search-documents";

const index: SearchIndex = {
  name: "products",
  fields: [
    { name: "id", type: "Edm.String", key: true },
    { name: "title", type: "Edm.String", searchable: true },
    { name: "description", type: "Edm.String", searchable: true },
    { name: "category", type: "Edm.String", filterable: true, facetable: true },
    {
      name: "embedding",
      type: "Collection(Edm.Single)",
      searchable: true,
      vectorSearchDimensions: 1536,
      vectorSearchProfileName: "vector-profile",
    },
  ],
  vectorSearch: {
    algorithms: [
      { name: "hnsw-algorithm", kind: "hnsw" },
    ],
    profiles: [
      { name: "vector-profile", algorithmConfigurationName: "hnsw-algorithm" },
    ],
  },
};

await indexClient.createOrUpdateIndex(index);

Indexer des documents

const documents = [
  { id: "1", title: "Widget", description: "A useful widget", category: "Tools", embedding: [...] },
  { id: "2", title: "Gadget", description: "A cool gadget", category: "Electronics", embedding: [...] },
];

const result = await searchClient.uploadDocuments(documents);
console.log(`Indexed ${result.results.length} documents`);

Recherche en texte intégral

const results = await searchClient.search("widget", {
  select: ["id", "title", "description"],
  filter: "category eq 'Tools'",
  orderBy: ["title asc"],
  top: 10,
});

for await (const result of results.results) {
  console.log(`${result.document.title}: ${result.score}`);
}

Recherche vectorielle

const queryVector = await getEmbedding("useful tool"); // Votre fonction d'embedding

const results = await searchClient.search("*", {
  vectorSearchOptions: {
    queries: [
      {
        kind: "vector",
        vector: queryVector,
        fields: ["embedding"],
        kNearestNeighborsCount: 10,
      },
    ],
  },
  select: ["id", "title", "description"],
});

for await (const result of results.results) {
  console.log(`${result.document.title}: ${result.score}`);
}

Recherche hybride (texte + vecteur)

const queryVector = await getEmbedding("useful tool");

const results = await searchClient.search("tool", {
  vectorSearchOptions: {
    queries: [
      {
        kind: "vector",
        vector: queryVector,
        fields: ["embedding"],
        kNearestNeighborsCount: 50,
      },
    ],
  },
  select: ["id", "title", "description"],
  top: 10,
});

Recherche sémantique

// L'index doit avoir une configuration sémantique
const index: SearchIndex = {
  name: "products",
  fields: [...],
  semanticSearch: {
    configurations: [
      {
        name: "semantic-config",
        prioritizedFields: {
          titleField: { name: "title" },
          contentFields: [{ name: "description" }],
        },
      },
    ],
  },
};

// Rechercher avec classement sémantique
const results = await searchClient.search("best tool for the job", {
  queryType: "semantic",
  semanticSearchOptions: {
    configurationName: "semantic-config",
    captions: { captionType: "extractive" },
    answers: { answerType: "extractive", count: 3 },
  },
  select: ["id", "title", "description"],
});

for await (const result of results.results) {
  console.log(`${result.document.title}`);
  console.log(`  Caption: ${result.captions?.[0]?.text}`);
  console.log(`  Reranker Score: ${result.rerankerScore}`);
}

Filtrage et facettes

// Syntaxe de filtre
const results = await searchClient.search("*", {
  filter: "category eq 'Electronics' and price lt 100",
  facets: ["category,count:10", "brand"],
});

// Accéder aux facettes
for (const [facetName, facetResults] of Object.entries(results.facets || {})) {
  console.log(`${facetName}:`);
  for (const facet of facetResults) {
    console.log(`  ${facet.value}: ${facet.count}`);
  }
}

Autocomplétion et suggestions

// Créer un suggesteur dans l'index
const index: SearchIndex = {
  name: "products",
  fields: [...],
  suggesters: [
    { name: "sg", sourceFields: ["title", "description"] },
  ],
};

// Autocomplétion
const autocomplete = await searchClient.autocomplete("wid", "sg", {
  mode: "twoTerms",
  top: 5,
});

// Suggestions
const suggestions = await searchClient.suggest("wid", "sg", {
  select: ["title"],
  top: 5,
});

Opérations par lot

// Upload, fusion, suppression par lot
const batch = [
  { upload: { id: "1", title: "New Item" } },
  { merge: { id: "2", title: "Updated Title" } },
  { delete: { id: "3" } },
];

const result = await searchClient.indexDocuments({ actions: batch });

Types principaux

import {
  SearchClient,
  SearchIndexClient,
  SearchIndexerClient,
  SearchIndex,
  SearchField,
  SearchOptions,
  VectorSearch,
  SemanticSearch,
  SearchIterator,
} from "@azure/search-documents";

Bonnes pratiques

1. Utilisez la recherche hybride - Combinez vecteur + texte pour les meilleurs résultats
2. Activez le classement sémantique - Améliore la pertinence pour les requêtes en langage naturel
3. Chargez les documents par lot - Utilisez uploadDocuments avec des tableaux, pas des documents individuels
4. Utilisez des filtres pour la sécurité - Implémentez la sécurité au niveau des documents avec des filtres
5. Indexez progressivement - Utilisez mergeOrUploadDocuments pour les mises à jour
6. Supervisez les performances des requêtes - Utilisez includeTotalCount: true avec parcimonie en production