RAGold

A web-based annotation tool for RAG (Retrieval-Augmented Generation) datasets. RAGold helps annotators create query-response pairs with relevant and distracting document chunks, following the Know Your RAG research framework.

Features

• Create and manage annotations with queries, responses, and associated document chunks.
• Classify queries by type: single-fact, summary, reasoning, or unanswerable.
• Upload source documents (any file type, max 10 MB each) with drag-and-drop and bulk upload support, and add optional notes.
• Link document chunks to specific uploaded documents.
• Export and import annotation datasets as zip archives containing both annotations and document files.
• Annotation metadata is stored in localStorage, document files in IndexedDB.
• Available in English and German.

Environment Variables

Variable	Description
VITE_CONTACT_INFO	Optional contact information displayed in the UI.

Data Format

Annotation datasets are exported as zip archives with the following structure:

ragold-<timestamp>.zip
├── annotations.json
└── files/
    └── <uuid>/
        └── <filename>

The annotations.json file contains:

{
  "version": 2,
  "author": "",
  "project": "",
  "notes": "",
  "language": "en",
  "createdAt": "2026-01-01T00:00:00.000Z",
  "updatedAt": "2026-01-01T00:00:00.000Z",
  "annotations": {
    "<uuid>": {
      "query": "",
      "queryType": "fact_single",
      "relevantChunks": [{ "content": "", "documentId": "<uuid>" }],
      "distractingChunks": [{ "content": "", "documentId": "<uuid>" }],
      "response": "",
      "notes": "",
      "createdAt": "2026-01-01T00:00:00.000Z",
      "updatedAt": "2026-01-01T00:00:00.000Z"
    }
  },
  "documents": {
    "<uuid>": {
      "name": "",
      "size": 0,
      "notes": "",
      "createdAt": "2026-01-01T00:00:00.000Z",
      "updatedAt": "2026-01-01T00:00:00.000Z"
    }
  }
}