WIBT tool

The multi-agent pipeline behind Wetenschap in begrijpelijke taal (WIBT), a research project by the KB Nationale Bibliotheek, the VU University Library, and the VU Computer Science department.

Give it a scientific paper (PDF) and it generates a Dutch, plain-language summary, tailored to one of three audiences — general public, GZ-psychologists, or information specialists at the Dutch parliament (Tweede Kamer) — while scoring the result on readability and factual accuracy against the source paper.

Project background, evaluation results, and publications: ubvu.github.io/wibt.

How it works

1. The pipeline generates several candidate English summaries of the paper.
2. Each candidate is scored on readability (syntactic clarity, jargon, information density, structural cohesion) and factuality — faithfulness (does every sentence follow from the paper?) and completeness (does the summary cover the paper's key facts?).
3. The highest-scoring summary is translated into Dutch through a chain of drafting, refinement, and proofreading agents.

Prompts are organized per audience (general, psychology, tweede-kamer) under prompts/, so tuning the tone or jargon rules for one audience doesn't require touching any code. See CLAUDE.md for the full pipeline architecture.

Getting started

Requires Python 3.14+.

uv sync                # or: pip install -e .

Create a .env file with your endpoint credentials and model settings — there's no .env.example yet, so see AppConfig.from_env() in src/config.py for the full list of required variables (API endpoint, default prompt context, and a model + temperature pair per agent).

Try the interactive demo (Marimo notebook):

marimo run src/gui/summary.py

Or run it from the command line:

python src/cli/summary.py -sc general -fc general -tc general -it 5 -i paper.pdf -st refine

Run any script in src/cli/ with --help for its full set of options.

Models

Built and evaluated with open models — gpt-oss-120b, Gemma3-12b, and TranslateGemma-12b — served via the VU Nebula AI infrastructure and the SURF AI-hub (WiLLMa). Any OpenAI-compatible endpoint works; set it through OPEN_AI_URL and OPEN_AI_TOKEN.

License

MIT — see LICENSE. Copyright University Library, Vrije Universiteit Amsterdam.