IntelligentHQ
IntelligentHQ
Skip to content
Cloud service Microsoft Foundry and AI services

Azure AI Search

Azure AI Search is a managed search and retrieval service that connects company content to apps, agents and language models. It supports full-text, vector, hybrid and multimodal queries and an agentic retrieval engine that powers Foundry IQ knowledge bases. It was formerly Azure Cognitive Search.1

Also known as Azure Cognitive Search, Foundry IQ

Intended for
Application and AI developers, Search and data engineers, Solution architects
Runs on
Azure public cloud regions (agentic retrieval and Serverless in select regions)
Official Microsoft product page (opens in a new tab) Official documentation (opens in a new tab)

At a glance

What is it?
Azure AI Search, formerly Azure Cognitive Search, is a fully managed Azure service for indexing and retrieving content. Each search service includes two engines: classic search, which answers one request against one index, and agentic retrieval, a multi-query pipeline in which a knowledge base plans subqueries, runs them in parallel across knowledge sources, reranks the results and returns content an agent can use. Learn lists Azure AI Search among the Foundry Tools and describes it as the service underneath Foundry IQ, the managed knowledge layer for agents in Microsoft Foundry; the Azure product page now presents the two names together.123
What does it do?
It ingests content by push (upload JSON) or pull (indexers for supported sources such as Azure Blob Storage, Azure Cosmos DB, SharePoint and OneLake), optionally enriching it with skills that chunk text and create embeddings. It runs keyword, vector, hybrid and multimodal queries with filters, facets, geo-search, synonyms and autocomplete, and can rerank with the semantic ranker. Agentic retrieval returns merged grounding content with optional references and an activity log. Document-level access control and Purview sensitivity labels can limit results to what each user may see.14
Who needs it?
Developers building retrieval-augmented generation, enterprise search or agents that must answer from company content with citations and user permissions. It suits teams that want control over indexing, ranking and security, such as Private Link, customer-managed keys and Microsoft Entra ID, rather than a closed search box.
What does it need?5
  • An Azure subscription and a search service on a Dedicated tier or the Serverless (preview) model
  • A supported region, especially for agentic retrieval
  • Content in a supported data source, or JSON documents you push
  • An Azure OpenAI model deployment if you use LLM query planning or answer synthesis
What it is not
Azure AI Search is not a system of record and indexes only JSON documents, so content is serialized first (indexers do this for supported sources). It is not a language model: agentic retrieval calls an Azure OpenAI model for query planning and answer synthesis, billed on the Azure OpenAI side. Vectors are optional; Microsoft notes that LLMs and agents do not require them. User-level permission inheritance works only for some sources (remote SharePoint, Blob Storage, ADLS Gen2); elsewhere you build security filters.15

Availability and licensing. Generally available. Agentic retrieval is GA through the 2026-04-01 REST API, with LLM query planning, answer synthesis and some knowledge sources in preview (2026-08-01-preview); the portals give preview-only access (Learn, checked 26 Oct 2026). Dedicated tiers bill by search units; the Serverless Developer tier is preview and started billing on 13 September 2026. The Free tier allows one service per subscription.12

The problem it solves

Language models do not know an organization's private documents, and pasting whole documents into prompts is slow, costly and ignores who is allowed to read what. Content also sits in many stores, each with its own query method.

Azure AI Search indexes or connects to those stores, returns the most relevant passages with references, and can enforce document permissions at query time, so apps and agents work from grounded context instead of guesses.

How it works

Classic search stores text in inverted indexes and vectors in vector indexes; indexing and querying are the same component in read-write and read-only modes.

Agentic retrieval adds a knowledge base on top:

  1. The app calls the knowledge base with a query and conversation history.
  2. At low or medium reasoning effort an LLM breaks the query into focused subqueries; at minimal effort this step is skipped.
  3. Subqueries run in parallel as keyword, vector or hybrid searches against indexed or remote knowledge sources, and each is semantically reranked.
  4. Results are merged and returned, with optional references and an activity log.

The 2026-04-01 REST API covers the generally available features; LLM query planning, answer synthesis and other preview features need the 2026-08-01-preview API. The Azure and Foundry portals give preview-only access.5

Azure AI Search architecture: components by layer and how they connectAgents & automationAI models & platformData & analyticsSecurity, identity &governanceAgent or application: Calls the knowledge base or index for groundingAgent or applicationKnowledge base (agentic retrieval): Plans subqueries and merges results from knowledge sourcesKnowledge base (agenticretrieval)Semantic ranker: Reranks results for relevanceSemantic rankerAzure OpenAI model: Optional query planning and answer synthesisAzure OpenAI modelData sources (Blob Storage, Cosmos DB, SharePoint, OneLake): Where company content livesData sources (Blob Storage,Cosmos DB, SharePoint,…Indexers and skillsets: Pull, chunk, vectorize and enrich contentIndexers and skillsetsSearch index (inverted and vector): Stores searchable text and vectorsSearch index (inverted andvector)Entra ID, Private Link and document ACLs: Controls who can query whatEntra ID, Private Link anddocument ACLs
Diagram as a list

  1. Agents & automation

    • Agent or applicationCalls the knowledge base or index for grounding

  2. AI models & platform

    • Knowledge base (agentic retrieval)Plans subqueries and merges results from knowledge sources
    • Semantic rankerReranks results for relevance
    • Azure OpenAI modelOptional query planning and answer synthesis

  3. Data & analytics

    • Data sources (Blob Storage, Cosmos DB, SharePoint, OneLake)Where company content livesConnects to Indexers and skillsets
    • Indexers and skillsetsPull, chunk, vectorize and enrich contentConnects to Search index (inverted and vector)
    • Search index (inverted and vector)Stores searchable text and vectors

  4. Security, identity & governance

    • Entra ID, Private Link and document ACLsControls who can query whatConnects to Search index (inverted and vector)
Components and connections as documented by Microsoft.

Capabilities

  • Full-text, vector, hybrid and multimodal search1

    Queries over indexed and remote content, with filters, facets, geo-search, synonyms and autocomplete.

    Why it matters: One service covers keyword precision and semantic similarity.

    Limits: Only JSON documents can be indexed; vectors need an embedding step.

  • Indexers and AI enrichment16

    Pull data from supported sources and use skills to chunk, vectorize and transform content during indexing.

    Why it matters: Reduces custom ingestion code for retrieval pipelines.

    Limits: Indexers have more limits on the Free tier, and long-running enrichment can time out on small services.

  • Agentic retrieval and knowledge bases15

    A knowledge base plans subqueries, runs them in parallel across knowledge sources, reranks and merges results for agents.

    Why it matters: Handles multi-part and context-dependent questions a single query misses.

    Limits: Adds latency and token-based cost; query planning, answer synthesis and some sources are preview; only select regions.

  • Semantic ranker56

    Reranks results for relevance (L2 reranking) and is used inside agentic retrieval.

    Why it matters: Improves result quality without changing the index.

    Limits: It is the billable part of the classic pipeline; it runs on Free but is not recommended there for large workloads.

  • Permission-aware retrieval4

    Document-level access control, ACL synchronization for supported sources, Purview sensitivity labels and queries under the caller's Entra identity.

    Why it matters: Agents and apps return only content each user is allowed to see.

    Limits: Automatic permission inheritance covers remote SharePoint, Blob Storage and ADLS Gen2; other sources need security filters.

  • Dedicated and Serverless pricing models6

    Dedicated tiers bill fixed capacity in search units; Serverless bills compute units and indexed storage.

    Why it matters: Choose predictable capacity or pay per use for bursty workloads.

    Limits: The Serverless Developer tier is preview, has no SLA, lacks some features and started billing on 13 September 2026.

  • Enterprise security16

    Microsoft Entra ID, Azure Private Link, customer-managed keys, IP firewall and availability zones.

    Why it matters: Meets common enterprise network and key management requirements.

    Limits: Private endpoints, customer-managed keys, IP firewall and availability zones are not available on the Free tier.

Practical use cases

Employees cannot find answers buried in policies and procedures stored in SharePoint and file shares.
Approach
Build a Foundry IQ knowledge base over those sources and connect it to an agent.
Role of Azure AI Search
Azure AI Search indexes or queries the content and enforces permissions.
Data, infrastructure and skills
Search service in a supported region, access to the sources and a Foundry project.
Type of benefit
Grounded answers
Caveats
Permission inheritance works only for supported sources; portal access to agentic retrieval is preview.
First step
Index one policy library and test ten real questions.
Website search returns poor matches for natural-language queries.
Approach
Use hybrid search with the semantic ranker and facets.
Role of Azure AI Search
Azure AI Search runs and ranks the queries.
Data, infrastructure and skills
A Dedicated tier service and product or content data in JSON.
Type of benefit
Search relevance
Caveats
Semantic ranking is billable; tune and measure before rollout.
First step
Compare keyword and hybrid results on a set of real queries.
An agent must answer complex questions that span several content sources.
Approach
Use agentic retrieval with a knowledge base pointing to several knowledge sources.
Role of Azure AI Search
The knowledge base plans subqueries, retrieves in parallel and merges results.
Data, infrastructure and skills
Supported region and, for planning, an Azure OpenAI deployment.
Type of benefit
Coverage of complex questions
Caveats
More latency and token-based cost than one query; LLM planning is preview.
First step
Start at minimal reasoning effort and raise it only if answers miss content.

Who uses it

Works with

Optional integration

Same family

  • Foundry ToolsLearn lists Azure AI Search among the Foundry Tools.

Optional integration for

Relationship labels follow Microsoft's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.

Getting started

  1. Choose a pricing model and tier

    Pick Dedicated (Basic or Standard for production) or Serverless (preview); use Free for tutorials.

    Check: You have one service created in a region that supports the features you need.

  2. Load content

    Use the import wizard or an indexer for a supported source, or push JSON documents.

    Check: The index shows the expected document count.

  3. Run a first query

    Follow the full-text or vector search quickstart.

    Check: Queries return relevant documents.

  4. Create a knowledge base

    Create knowledge sources and a knowledge base, or use Foundry IQ in the Foundry portal.

    Check: A retrieve call returns merged content with references.

  5. Connect an agent

    Connect the knowledge base to a Foundry Agent Service agent.

    Check: The agent answers with citations to your documents.

  6. Pin production to the GA API

    Use the 2026-04-01 REST API for production; keep preview features out of production paths.

    Check: Production code calls only the GA API version.

Official resources

Deciding whether you need it?

Three free tools apply Microsoft's own published guidance to your situation, and say when Azure AI Search or any new technology is not needed.

Sources

Each statement above links to the source it comes from. Labels say who reported it.

  1. Introduction to Azure AI Search (opens in a new tab) Microsoft · Vendor-reported · link checked 27 Oct 2026
  2. Azure AI Search product page (opens in a new tab) Microsoft · Vendor-reported · link checked 27 Oct 2026
  3. What are Foundry Tools? (opens in a new tab) Microsoft · Vendor-reported · link checked 27 Oct 2026
  4. What is Foundry IQ? (opens in a new tab) Microsoft · Vendor-reported
  5. Agentic retrieval overview (Azure AI Search) (opens in a new tab) Microsoft · Vendor-reported
  6. Choose a pricing model and service tier (Azure AI Search) (opens in a new tab) Microsoft · Vendor-reported

Fill out the form below to request your copy.

Name(Required)