Skip to main content
A Retrieval preset is a named bundle of retrieval settings. Instead of tuning every Agent by hand, you tune once and reuse.

​
Anatomy of a preset

  • Name — how you’ll refer to it.
  • Top K — how many chunks to return per query. Default 6.
  • Score cutoff — minimum fused score to include. Default 0.15 (RRF space).
  • Vector weight / Keyword weight — hybrid balance.
  • Reranker — optional cross-encoder rerank pass on the top 20 candidates. Improves quality; adds ~200ms latency.
  • Query expansion — on/off, count.
  • Filters — always-on metadata filters.
  • Grounding — verify against causal graph, on/off.

​
Creating a preset

  1. Workspace → Retrieval presets → New.
  2. Set the knobs.
  3. Test with a few real queries.
  4. Save.
Attach to an Agent from the Agent block’s Inspector, Retrieval preset field.

​
Common presets to start from

Nora ships three built-in presets you can duplicate and modify:
  • Balanced (default) — hybrid, top 6, no reranker. Good general starting point.
  • Precise — hybrid + reranker + higher cutoff. For high-stakes answers where quality > latency.
  • Fast — keyword-only, top 4, no reranker. For live chat where sub-second response matters.

​
Per-Agent overrides

Sometimes an Agent needs a slight variation. You can override individual preset fields on the Agent without cloning the whole preset — the Inspector shows an “override” indicator when a field diverges from the preset.

​
Preset versions

Presets are versioned like Flows. Changing one doesn’t affect production runs until you publish the Agent that uses it. This means you can safely tune a preset in Draft and A/B test the difference.

​
Testing multiple presets side-by-side

In the preset editor, hit Compare. Pick 2-3 presets and a query. Nora runs the query against each and shows the ranked results side by side with score breakdowns. Great for choosing between a “faster but noisier” and “slower but tighter” configuration.