📋 Reference · Last updated: September 17, 2026
<aside>
🎛️
WHAT IT DOES
Every setting in a RAG project — what it does and when to change it — grouped by the five tabs you see when you create or edit one: General, Retrieval, Generation, Groundedness, and Experience. Sensible defaults are set for you; most projects only need a few of these touched. Tune based on what the RAG Projects test page shows.
</aside>
General
What the project is and what it answers with.
| Setting |
What it does / when to change |
| Name, Description |
Label the project so collaborators know its purpose. |
| Knowledge base |
The library this project searches. Required — without one, runs fail. See Knowledge Bases. |
| AI model |
Which model writes the answer. Leave on Workspace default unless you need a specific model. |
| Persona / system prompt |
Sets tone and role (e.g. "a concise, friendly city-services assistant"). Optional. |
Retrieval — finding the right content
How the project searches your knowledge base for each question.
- Query type — Simple (keywords), Semantic (meaning & context), Vector (embedding similarity), or the combined Vector + Simple / Vector + Semantic. Vector modes need a knowledge base with vector search turned on — if it isn't, the project stays selectable but quietly searches text only, so check the knowledge base first. Change it if the project misses obviously-relevant content — a vector or semantic mode handles natural-language questions far better than keywords. How the content was split doesn't restrict this choice; see Choosing a split style.
- Search filter mode — Disabled, Hybrid, or Strict. Controls how custom-field filters narrow results. Leave Disabled unless you filter on custom fields.
- Retrieved documents (Top-N) — how many passages the model sees per question (default 5, up to 50). More gives the model additional context but can add noise; raise it if answers miss detail, lower it if they wander.
- Extract relevant sentences — passes only the relevant lines from each passage instead of the whole thing. Helps when passages are long.
- Query optimization steps — pre-search steps that sharpen the question: Classify intent, Generate direct answer, Analyze specificity & ambiguity (asks a clarifying question when a query is vague — multi-turn, so it's skipped in the test page), and Refine query & extract entities. Turn off any you don't need.
- Clean search results — strips unrelated text from results before the answer is written. Useful when passages are noisy.
Generation — writing the answer
How the model composes the answer once it has the content. The defaults produce focused, factual answers.
| Setting |
Default |
What it does |
| Temperature |
0.3 |
0 = deterministic, 1 = balanced, 2 = creative. Keep low for factual answers. |
| Top P |
1 |
Nucleus sampling (0–1). Leave at 1 unless you know you need it. |
| Max response tokens |
4096 |
Caps answer length. Raise for long-form answers. |
| Frequency penalty |
0 |
Higher discourages repeating the same words. |
| Presence penalty |
0 |
Higher encourages introducing new topics. |
Groundedness — staying accurate
How tightly the answer must stick to your content — the most important group for trustworthy answers.
- Limit responses to knowledge base content — on by default. The agent answers only from your material and says so when something isn't covered, instead of guessing. Set the out-of-scope reply it uses.
- Groundedness strictness — Low (minor inferences allowed), Medium (close match, slight rewording), or High (must adhere to sources). Keep High for public-facing, factual agents.