🛠️ How-to guide

<aside> 🧠

WHAT IT DOES

A RAG project is the answering recipe your agent uses. It points at a knowledge base and controls how questions are searched, how answers are written, and how tightly the agent sticks to your content. (RAG is glossed in the Glossary; it's sometimes called a knowledge project.)

</aside>

RAG project vs. knowledge base

Think of it as two layers. A Knowledge Bases is the content and its search index — the what. A RAG project is how you search and answer over that content — the how. One knowledge base can sit behind several RAG projects, each tuned differently. A RAG project with no knowledge base attached can't run yet.

The RAG Projects list: each project shows its Status — KB bound, No KB, or Legacy — its Top-N, and when it was last updated. Filter by All / KB bound / No KB, or start one with New project.

Creating a RAG project

RAG Projects is its own area in the sidebar. Select New project. The setup is organized into five tabs. For what every setting means and when to change it, see RAG project settings explained; to validate a project before connecting it, see Testing a RAG project.

General

• Name and Description.
• Knowledge base — the library this project searches (required for it to run).
• AI model — which model writes the answers, or the workspace default.
• Persona — an optional system prompt that sets tone and role.

Retrieval — how your content is searched

• Query type — Simple (keywords), Semantic (meaning and context), Vector (embedding similarity), or the combined Vector + Simple / Vector + Semantic. Vector modes need a knowledge base with vector search on.
• Search filter mode — Disabled, Hybrid, or Strict control on custom-field filtering.
• Retrieved documents — how many passages the model sees per question (default 5, up to 50).
• Extract relevant sentences — pass only the relevant lines from each passage.
• Query optimization steps — pre-search steps that sharpen the question (classify intent, try a direct answer, spot vague questions, refine wording and pull out key terms).
• Clean search results — strip unrelated text before answering.

Generation — how answers are written