🛠️ How-to guide

<aside> 🧪

WHAT IT DOES

Before you connect a RAG project to an agent, the test page lets you ask real questions and see exactly how it answers — the answer, the sources and passages it used, the tokens it spent, and a step-by-step trace of the pipeline. It's how you tune a project with confidence, without touching a live agent.

</aside>

Opening the test page

From RAG Projects, select a project to open its test page. If the project isn't bound to a knowledge base, you'll see a warning — runs will fail until you select Edit and pick one on the General tab.

Asking a question

Type a question in the ask bar at the top — "Ask something the knowledge base could answer…" — and press Enter. The project runs its full pipeline against the bound knowledge base and the result appears below.

The RAG project test page after a run — the settings used, a pipeline trace of each step with timings, the answer with a citation, and token usage.

Reading the result

The result is laid out top to bottom:

• Configuration summary — cards showing the AI model, Query type and Search filter mode, Top-N documents, Temperature, and Persona this run used, so you always know which settings produced the answer.
• Answer — the generated answer. Inline citations link to the exact passages behind each claim.
• Sources cited — the sources the answer drew on; hover one to highlight the matching passage below.
• Retrieved chunks — the actual passages the search returned, each with a relevance score and its source. This is where you confirm the right material was found — if a good answer's passages look wrong, your retrieval settings need work.
• Token usage — prompt, completion, and total tokens for the run, with an approximate cost estimate (for reference — you're not billed per test).

The pipeline trace

Above the answer, a timeline shows every step the project ran — typically Classify intent, Generate direct answer, Refine query, Search, Clean search results, Generate answer, and Groundedness check — with how long each took. Expand any step to see its Request and Response. Use the trace to see where time goes and to understand how a step (like the groundedness check) shaped the final answer.

Using the results to tune the project

The test page is a feedback loop: spot a problem, change the matching setting in Edit, and re-run.

• Missed relevant content? Try a Vector or Semantic Query type, or raise Top-N.
• Drifted off-source? Raise Groundedness strictness and keep responses limited to the knowledge base.