AgentStack Docs

Configure retrieval

Choose how an agent searches its knowledge, tune confidence guidance, and test the result against real questions.

Open Sources → Retrieval to configure how an agent searches its selected knowledge. Use the Save Retrieval Settings action to save your changes.

The default mode is Standard, with a 30% Confidence Threshold. The default search model is Gemini 3.5 Flash, with a maximum of 10 documents for deep search.

Set the confidence threshold

The threshold ranges from 0% to 100%. It is a guidance threshold, not a refusal switch. When retrieval returns no result, or when the score is below the threshold, AgentStack guides the agent to say that it does not always have enough information, and to offer more help.

Raise the threshold only when low-confidence answers are not helpful. After you change the threshold, test it with known questions and edge-case questions.

Choose a mode

Standard mode uses fast embedding-based search. It is available on every plan.

Hybrid mode runs Standard first. It uses a deeper step when confidence is low. Hybrid is available on Standard and higher plans.

Deep Search reads full documents and follows cross-references.

Auto classifies each question. It currently chooses Standard mode or Hybrid mode.

Deep Search and Auto are available on Pro and Enterprise plans.

Hybrid runs its deeper step only when Standard confidence is 80% or lower. If deep search fails, Hybrid keeps its Standard result. If Deep Search mode fails, it also falls back to Standard. Auto does not choose Deep Search on its own. Treat Auto as a choice between Standard mode and Hybrid mode only.

Deep-search controls

When the plan allows a mode other than Standard, the page shows Search Model and Max Documents. The Max Documents count ranges from 1 to 50.

The Search Model picker lists several provider models. The page does not state a plan or cost for each model. Test a model with a controlled evaluation before you choose it.

Save your changes. Then compare the same prompts in Playground.

Retrieval tuning cannot correct a source that is missing, outdated, or contradictory. Check Sources first.

On this page