Azure AI search deep dive: From SQL queries to AI-powered retrieval

This article is the first part of the Search Trilogy, a series comparing three major enterprise search platforms: Azure AI Search, Apache Solr, and Elasticsearch.

In this installment, we focus on Azure AI Search, Microsoft’s fully managed search service that combines traditional inverted-index search with AI enrichment and vector-based retrieval.

Limitations of traditional SQL search

SQL-based search is fundamentally limited to exact or pattern-based matching. It lacks:

• Relevance ranking based on scoring models
• Linguistic processing (stemming, synonyms)
• Autocomplete and suggestion capabilities
• Semantic understanding of queries

As a result, SQL queries return correct results but not necessarily relevant ones, especially for large or unstructured datasets.

Azure AI search architecture overview

Azure AI Search introduces a decoupled architecture where search is performed on a precomputed index rather than directly on the data source.

The system acts as an intermediary between:

• External data stores (unindexed raw data)
• Client applications (query execution and result consumption)

This separation enables optimized query performance, advanced ranking, and AI-driven enrichment.

Core components

Data source

Data source defines the origin of the data (Azure SQL, Blob Storage, Cosmos DB, etc.).Important: Azure AI Search does not query the data source at runtime—all queries are executed against the index.

Index

The central data structure is similar to a search-optimized schema.

Each index:

• Has a defined schema
• Contains fields with attributes such as:
  • searchable
  • filterable
  • sortable
  • facetable
  • retrievable

Proper field configuration is critical for query performance and flexibility.

Document

Represents the atomic unit of search.

• Equivalent to a JSON object
• Contains fields defined in the index schema
• Must include a unique key

Documents are ingested into the index via push or pull mechanisms.

{
  "productId": "p001",
  "name": "Wireless Bluetooth Headphones",
  "description": "High-quality noise-cancelling headphones with long battery life.",
  "price": 199.99,
  "category": ["Audio", "Headphones"],
  "rating": 4.5
}

Indexer (pull model)

A scheduled crawler that:

• Extracts data from the data source
• Maps fields to the index schema
• Populates the index

Each indexer maintains a 1:1 relationship with a data source and index.

Push model (custom ingestion)

Data is explicitly pushed to the index via:

• REST API
• SDKs (Java, Python, C#, etc.)

This approach allows integration with non-Azure data sources or custom pipelines.

{
  "name": "(required) String that uniquely identifies the indexer",
  "description": "(optional)",
  "dataSourceName": "(required) String indicating which existing data source to use",
  "targetIndexName": "(required) String indicating which existing index to use",
  "parameters": {
    "batchSize": null,
    "maxFailedItems": 0,
    "maxFailedItemsPerBatch": 0,
    "base64EncodeKeys": false,
    "configuration": {}
  },
  "fieldMappings": "(optional) unless field discrepancies need resolution",
  "disabled": null,
  "schedule": null,
  "encryptionKey": null
}

Skillsets (AI enrichment pipeline)

Skillsets define a pipeline of cognitive transformations applied during indexing.

Common capabilities:

• OCR for text extraction from images/PDFs
• Language detection and translation
• Key phrase extraction
• Entity recognition

Skillsets operate before data is indexed, enriching documents with additional searchable metadata.

Query layer

Azure AI Search supports multiple query paradigms:

• Full-text search (Lucene-based, fuzzy, autocomplete)
• Vector search (embedding-based similarity search)
• Hybrid search (combining lexical + vector queries)

Query execution includes:

• Scoring and ranking
• Filtering ($filter)
• Sorting ($orderby)
• Faceting

{
  "method": "POST",
  "url": "/indexes/products/docs/search?api-version=2023-07-01",
  "body": {
    "search": "wireless headphones",
    "filter": "brand eq 'Sony'",
    "top": 5
  }
}

Infrastructure as code considerations

Azure AI Search has partial Terraform support:

Supported:

azurerm_search_service (search service provisioning)

Not supported:

• Indexes
• Indexers
• Skillsets

These must be created via:

• REST API
• Azure SDKs

This introduces a hybrid provisioning model in which IaC handles infrastructure, while application-level configuration is API-driven.

resource "azurerm_search_service" "example" {
  name                = "my-search-service"
  resource_group_name = "rg-example"
  location            = "West Europe"
  sku                 = "standard"
}

Retrieval-augmented generation (RAG)

Azure AI Search integrates directly into RAG architectures.

Typical flow:

1. User submits a natural language query
2. Search service retrieves relevant documents (keyword, vector, or hybrid)
3. Results are passed to an LLM (e.g., Azure OpenAI)
4. The model generates a grounded response

This approach improves:

• Accuracy
• Context awareness
• Explainability of AI outputs

Practical use cases

Azure AI Search is particularly effective in scenarios involving:

• Large-scale document search (PDFs, knowledge bases)
• E-commerce product discovery (faceted filtering, ranking)
• AI assistants with contextual retrieval
• Enterprise search across heterogeneous data sources

Conclusion

Azure AI Search provides a search-first architecture that separates data storage from retrieval, enabling scalable, performant, and intelligent search experiences.

Key strengths:

• Precomputed indexing for performance
• AI enrichment for unstructured data
• Hybrid search (lexical + vector)
• Integration with generative AI (RAG)

In the next part of the Search Trilogy, we will analyze Apache Solr, focusing on its flexibility, configuration model, and on-premise capabilities.