• +49-(0)721-402485-12
Ihre Experten für XML, XQuery und XML-Datenbanken

StrucRAG – Methodical framework for measuring and reducing information loss in structured RAG pipelines

What is the goal of the project?

The goal of StrucRAG is the development of a methodical framework for measuring and reducing information loss that occurs during the processing of structured documents for RAG (Retrieval-Augmented Generation) systems. This includes the creation of a referenced corpus of structured documents, the development of suitable quality metrics, as well as the design of layout- and semantics-aware representation and chunking methods. The results will serve as a scientific and technical foundation for future product developments.

Project duration: 05.01.2026 to 22.12.2027

What are the three biggest risks?

  • It is unclear which structure and layout features, individually and in combination, measurably contribute to quality improvement and whether their influence can be separated from model and prompt effects. There is a risk that no stable, cross-domain usable metrics and benchmarks can be defined.
  • The developed representation and chunking methods might not show statistically significant advantages over baselines despite additional effort.
  • The complexity of layout analysis could lead to inconsistent annotations and unreliable results. Improvements might not be robustly transferable to other document domains such as legal texts or technical manuals. A valid outcome could also be to demonstrate the limits of modelability and to formulate criteria for when structured processing does not provide added value.

What will be the outcome of the project?

The project forms the basis for the development of a framework with a benchmark corpus, metrics, as well as layout- and semantics-aware methods for the measurable reduction of information loss. Reference corpora, indicators for measuring structure, retrieval, and answer quality, as well as new representation and chunking methods will be designed and experimentally tested in prototype pipelines. The results will be documented in reproducible protocols and serve as a foundation for future product developments.