Skip to main content
A retrieval preset is a named bundle of retrieval settings. Instead of tuning from scratch every time, pick a well-crafted preset and apply it in one click. The chip next to the retrieval page title shows the active preset.
Preset menu showing New preset, Duplicate current, and From starter preset options

​
What a preset holds

Every knob in the retrieval settings dock is part of the preset.
  • Models: HyDE chat model, embedding model
  • Search: method (vector · keyword · hybrid · KG-rooted), Top k (default 8)
  • Query: HyDE on/off
  • Ranking: vector/keyword weights, fusion (Weighted/RRF), minimum similarity
  • Rerank: provider (off · Cohere · Jina)
  • Causal reasoning: Reason over the graph plus its sub-options

​
Creating a preset

Click the preset chip next to the title. The menu offers:
  • New preset: start blank.
  • Duplicate current: copy the active preset to fork a variant.
  • From starter preset: start from a purpose-built template.
  • Refinery settings: open the retrieval settings dock.

​
Starter templates

Seven built-in templates cover the common shapes. Create a preset from one and tweak only what you need.
  • Auto
  • Fast & broad
  • Precision (RAG-grade)
  • KG-anchored
  • Multi-hop relations
  • Graph reasoning
  • Exact
They fall into speed-first (Fast & broad), precision-first (Precision, Exact), and causal-graph-driven (KG-anchored, Multi-hop relations, Graph reasoning) buckets.

​
Testing presets side by side

Flip the top toggle to Compare mode.
  • Single query: pick up to 4 presets and run one query against each. Ranked results appear side by side. Useful for judging fast-but-fuzzy vs. slow-but-precise on real output.
  • Eval set: run a saved batch of queries at once. See Diagnostics.