Plan and manage an Azure AI solution 1.1 Choose the appropriate Foundry services for generative AI and agents
A Foundry IQ knowledge base has an LLM connection for query planning.
Study guide skill: Choose an appropriate method for retrieval and indexing
Question B2-002
A Foundry IQ knowledge base has an LLM connection for query planning. Users ask multi-part questions, and when the first round of results isn't relevant enough, the engine must automatically run a follow-up search. Which retrieval reasoning effort (preview) should you configure?
Select one answer.
Answer: B
Iterative search depends on the medium retrieval reasoning effort: after the first search, a semantic classifier evaluates the retrieved documents and, if they aren't sufficiently relevant, the service performs a follow-up iteration with a revised query plan. Low runs a single pass of LLM query planning and source selection without iterating, and minimal disables LLM query planning entirely. There is no high level; besides minimal, low, and medium, the newer auto level starts with a lightweight pass and continues with LLM-based query planning only up to medium effort when needed, so it doesn't guarantee the follow-up iteration. Retrieval reasoning effort is currently a preview capability of agentic retrieval.
Deep dive
How it works
Retrieval reasoning effort (preview) sets how much LLM processing agentic retrieval does for query planning. More effort can improve relevance, but it also takes longer and uses billable LLM resources. You can set it on the knowledge base as the default, or per retrieve request to override it. If neither is set, the service uses low.
The levels on the page:
- minimal: Disables LLM-based query planning. It issues direct text and vector searches across the knowledge sources and returns the best-matching passages.
- low: The default. It runs a single pass of LLM-based query planning and knowledge source selection.
- medium: Adds deeper search. After the first search, a high-precision semantic classifier evaluates the retrieved documents. If the initial results aren't sufficiently relevant, the service performs one follow-up iteration using a revised query plan.
- auto: Starts with a lightweight pass and continues with LLM-based query planning, up to medium effort, only when the first pass doesn't provide enough grounding. It needs a newer preview REST API version.
The requirement is "automatically run a follow-up search when the first round isn't relevant enough". That's the definition of medium's iterative search.
The other options, and when you'd use them
- High: There's no high level. The levels listed are minimal, low, medium and auto.
- Minimal: Minimal disables LLM-based query planning entirely. It suits migrations from the Search API or cases where you manage query planning yourself.
- Low: Low runs a single pass of query planning and source selection. It's a balance between latency and depth, but it doesn't iterate.
Exam tip
"Follow-up iteration when results aren't relevant" = medium. "Single planning pass" = low (the default). "No LLM planning" = minimal.
References
Microsoft Learn references
Last reviewed: