Settings

Switch the active model and RAG instance, and edit the harness answer system prompt. Changes apply to the next call — no restart.

Model

The Qwen model every harness/answer call uses. Source: live Ollama /api/tags.

Active Model

Switch applies to the next LLM call; larger models are slower. A restart reverts to the env default.

RAG Instance

Which named corpus the answerer retrieves from. Pick No RAG to answer from the bare model.

Active RAG Instance

Deleting removes the selected instance's corpus permanently. The example default instance and No RAG can't be deleted.

Retrieval Features

Toggle the advanced RAG methods live — no restart. Every answer's metadata shows which were active, so you can A/B with and without. Guides: examples/feature-guides/.

Hybrid retrieval (vector + BM25) Cross-encoder reranker Grounding reinjection

Query rewrite mode

Top-K (chunks injected) Fetch-K (reranker first pass)

System Prompt

The system prompt for the harness answer path only. It does not affect the judge or any other code-defined prompt.

Saved Prompts

Harness Answer System Prompt

Quick-Save Current Prompt as a Preset