Implement generative AI and agentic solutions 2.1 Build generative applications by using Foundry
An agent queries a Foundry IQ knowledge base built on Azure Blob Storage and an existing…
Study guide skill: Implement retrieval-augmented generation (RAG) in an application
Question B2-014
An agent queries a Foundry IQ knowledge base built on Azure Blob Storage and an existing search index (no web knowledge source). The agent should reason over the retrieved content itself and compose the final answer for the user. Which output should the knowledge base return?
Select one answer.
Answer: B
For most Foundry IQ scenarios, extractive data is recommended: it returns raw content that the agent can reason over and incorporate into its own response. Answer synthesis (preview) is reserved for standalone apps where retrieval output goes straight to users without agent processing (it's required only for web knowledge sources). An activity log or titles alone don't give the agent grounding content.
Deep dive
How it works
A Foundry IQ knowledge base can return one of two kinds of output, and the FAQ gives a clear default:
- Extractive data: raw retrieved content that agents can reason over and incorporate into their own responses. The FAQ says to use this for most Foundry IQ scenarios.
- Answer synthesis (preview): an LLM writes a complete, natural-language answer from the retrieved content. Reserve it for standalone apps where retrieval output goes straight to users without agent processing. It's required only for web knowledge sources.
This agent's job is to reason over the content and compose the final answer itself, and its sources are Blob Storage and an existing search index, with no web source. That's the extractive data case. If the knowledge base synthesized an answer first, the agent would be reasoning over a summary rather than the evidence, and you'd pay for an extra LLM step.
Extractive output also keeps the retrieved references, so the agent can cite sources in its reply.
The other options, and when you'd use them
- Only document titles: Titles carry no grounding content. The agent needs the raw content that extractive data returns.
- Only the query plan activity log: The activity log describes what the engine did, such as query planning and synthesis steps. It isn't the retrieved content.
- Answer synthesis (preview): It's reserved for standalone apps where the output goes directly to users without agent processing. It's required only for web knowledge sources, which this knowledge base doesn't have.
Exam tip
Agent composes the answer = extractive data. Standalone app showing retrieval output directly (or any web knowledge source) = answer synthesis.
References
Last reviewed: