What "Knowledge base" mode does

  • Second generation option under Datasets → Generate dataset: generates both inputs and outputs from a source document, producing a ground truth dataset. 3:13
  • Accepted source formats: PDF, CSV, or Markdown; a short document is recommended to keep generation quick. 3:13

Steps demonstrated

  1. Use a PDF of the European AI Act as the source document. 3:30
  2. Drop the document, name the dataset, and request 10 questions. 3:30

How it works and why it matters

  • In the background the platform processes the document, extracts context, and generates questions to it; because answers come directly from the trusted source, they can be expected to be reliable (with optional manual review/tweaking). 4:09
  • Intended RAG evaluation workflow: take the question part, pass it through your system to generate answers, and compare those answers against the ground truth. 4:09
flowchart LR
    A[Source document<br/>PDF / CSV / MD] --> B[Platform extracts context]
    B --> C[Generated Q&A<br/>ground truth dataset]
    C -->|questions only| D[Your RAG system]
    D --> E[Generated answers]
    C -->|ground truth answers| F[Compare]
    E --> F