A Foundry IQ knowledge base uses low retrieval reasoning effort (preview) across several…

How it works

In a Foundry IQ knowledge base, the retrieval reasoning effort decides how much LLM processing each query gets. At low and medium effort, the LLM plans and selects which knowledge sources to query. Medium adds an iterative pass for deeper results. At minimal effort, LLM processing is bypassed.

So when the LLM skips the HR source, the fix is to give it better signals. Microsoft lists what informs source selection at low and medium effort:

1. The name of the knowledge source.
2. The description of the index (for indexed knowledge sources).
3. The retrievalInstructions in the knowledge base definition or the retrieve action.

Retrieval instructions guide which knowledge sources the LLM selects or skips, and they work like a prompt. A clear name ("hr-policies"), a description that says what the index holds, and an instruction such as "use the HR policies source for questions about leave, benefits or conduct" let the LLM route HR questions there without querying it every time.

The other options, and when you'd use them

• Minimal effort: Minimal bypasses LLM processing. It doesn't make the LLM plan more carefully.
• alwaysQuery = true: It includes the source in every query, regardless of reasoning effort, which the requirement rules out.
• Move to a remote source: Index descriptions are a selection signal only for indexed sources, and remote sources have no backing index. Moving would remove a signal, not add one.

Exam tip

LLM skips the right source = improve source name, index description and retrievalInstructions. alwaysQuery = every query. Minimal = no LLM.

References

• What is a knowledge source?