We’re extending the agentic retrieval API with answer synthesis and knowledge sources so you can ground agents across multiple indexes and generate answers with citations.

What are knowledge sources?

Knowledge sources in Azure AI Search provide grounding content for agents. You can create knowledge sources using existing search indexes or Azure Blob Storage containers. Blob knowledge sources use integrated vectorization to generate a complete indexing pipeline that automatically chunks and vectorizes multimodal content, then stores it inside a new search index.

Knowledge sources are defined once, then can be used by many different agents for grounding.

What is knowledge source selection?

Agentic retrieval analyses the entire chat thread using an LLM to determine which knowledge sources are most relevant to the underlying information need of the conversation using their names and descriptions. You can provide natural language retrieval instructions to influence how agentic retrieval considers your knowledge sources during selection.  The LLM query planner then incorporates these knowledge sources into a focused query plan to retrieve the most relevant content for your agent.

What is answer synthesis?

Agentic retrieval now can generate complete answers to user questions using an LLM. These answers are grounded in retrieved content from your knowledge sources and include inline citations. You can provide natural language answer instructions to provide additional guidance around the style and tone of the generated answer.

These improvements are only available through the agentic retrieval API. You can get started with these updates by reading the agentic retrieval documentation and trying out the agentic retrieval sample.