Configure data sources for knowledge bases
Unit 4 of 7 · Module 10: Build knowledge-enhanced AI agents with Foundry IQ
Go deeper
Unlocks when you've watched the lesson.
Read this lesson
Where the knowledge lives
A Foundry IQ knowledge base doesn't hold knowledge by itself. It's built from knowledge sources: connections to the places your content actually lives. Think of a restaurant kitchen during a busy evening service. Some ingredients are already in the pantry, washed, chopped, and labeled before the first order comes in. Others are ordered from a supplier the moment a dish needs them. Knowledge sources work the same way: an indexed source is copied in and prepared ahead of time, and a remote source is asked live, at the moment a question arrives. One knowledge base can mix both kinds, and when an agent asks a question, the retrieval engine queries all of them in one go and ranks every result together.
Indexed sources: the prepped pantry
Indexed sources are the pantry. If you already have an Azure AI Search index, a search index source simply wraps it. For documents in an Azure Blob Storage container, or data in a Microsoft OneLake lakehouse, Foundry IQ builds the whole preparation line for you: a data source, a skillset, an indexer, and an index. Along that line, each document is split into chunks and turned into embeddings, like chopping vegetables and labeling the jars. There are preview options too: indexing a SharePoint site, uploading files straight into the service, or pulling rows from an Azure SQL table. Because everything sits in one prepared index, queries run fast, with keyword, vector, or hybrid search, and answers can link back to the exact document. The catch is freshness, so indexer runs on a schedule restock the pantry as your documents change.
Remote sources: call the supplier
Remote sources are the supplier on the phone. Nothing is copied into Azure AI Search: the retrieval engine asks the other system through its own API, right when the question comes in. The web source asks Bing, for public information that changes every day. A remote SharePoint source, in preview, searches SharePoint directly, and SharePoint checks what each user is allowed to see. Other preview options reach live results from an MCP server, a Microsoft Fabric data agent, or Work IQ. The upside is freshness, with no copy to store or keep in sync. The trade-off: each answer waits on that system, depends on it being available, and doesn't come with the citation links an index can give.
Choosing the right source
So how do you choose? Start from where the content already lives, and what the answers need. Product manuals sitting as PDFs in Blob Storage: a blob source, prepared ahead, fast and citable. An index your search team already tuned: wrap it with a search index source, instead of building a second one. Company files in SharePoint, where each person may only see their own team's documents: remote SharePoint keeps the permissions in SharePoint's hands, while indexed SharePoint trades a little freshness for speed. Sales figures in a Fabric lakehouse: a OneLake source. And news, prices, or anything public that changes by the hour: the web source. One practical rule: create the knowledge sources first, on the same search service, and then build the knowledge base that points at them.
Permissions and tidy data
Two last habits keep the kitchen safe and the food good. First, permissions: Foundry IQ can carry document access lists and Microsoft Purview sensitivity labels into the index, and check them at query time against the person who's asking, so an agent only returns what that user is allowed to read. Second, the quality of your data shapes every answer. Clean, current documents retrieve better than messy duplicates. Give each knowledge source a clear name and description, because the retrieval planner reads them to decide which sources to ask. And if one source must be checked for every single question, you can mark it to always be queried.
Facts from Microsoft's documentation: What is a Knowledge Source?; What is Foundry IQ? (CC BY 4.0), rewritten in Loop's own words and examples. Narration, animation, and quiz by Loop. Microsoft, Azure, and Microsoft Foundry are trademarks of Microsoft. Loop is independent and not affiliated with or endorsed by Microsoft.