[research]
We publish what we ship, and ship what we publish.
Papers, whitepapers, and posts from the people who run the audits and build the systems. Our research focuses on AI systems that can show their work: grounded retrieval, attribution, hallucination detection, rule enforcement, and audit evidence.
Start here
A short selection from the research library while the full archive grows.
-
The verbatim pipeline constrains answer content through extraction + templating.
-
Token-level hallucination detection for RAG, trained on RAGTruth.
-
Task-conditioned pruning for noisy coding-agent tool output.
Published work
Papers, whitepapers, and posts, listed together because the argument, code, and deployments inform each other.
-
VerbatimRAG is applied to 110K+ papers of the ACL Anthology
-
Squeez is a benchmark, model, and CLI for reducing noisy coding-agent tool output to the smallest verbatim evidence block worth keeping.
-
Task-conditioned pruning for noisy coding-agent tool output.
-
A practical introduction to VerbatimRAG, the KR Labs approach to retrieval where answers are assembled from source spans instead of generated freely from retrieved context.
-
TinyLettuce compresses the LettuceDetect approach into smaller encoder models for deployments where latency, memory, and operating cost matter.
-
The verbatim pipeline constrains answer content through extraction + templating.
-
The multilingual LettuceDetect release extends token-level RAG hallucination detection to EuroBERT-backed models for major European languages.
-
An introduction to LettuceDetect as a token-level hallucination detection framework for retrieval-augmented generation systems.
-
Token-level hallucination detection for RAG, trained on RAGTruth.
-
Open-source framework for explainable information extraction.
No published work matches this filter.
Student research
Theses supervised by KR Labs researchers and university collaborators. They extend the same research lines into retrieval, hallucination detection, rule learning, and explainable information extraction.
2026
-
Retrieval Augmented Generation: A Multi-Stage Architecture for Verbatim Financial Question Answering
Selenge, supervised by Ádám Kovács
-
Latency-Tiered Hallucination Detection: Optimizing Supervised-Unsupervised Pipelines for RAG Systems
Rathmayr, supervised by Ádám Kovács
-
Real-time Prevention of Factual Hallucinations in Retrieval-Augmented Generation
Beccard, supervised by Ádám Kovács
-
Evaluating Extraction-Based RAG: A Systematic Assessment of VerbatimRAG on the CLAPNQ Benchmark
Kunerth, supervised by Gábor Recski
2025
-
Symbolic natural language inference for German open information extraction
Ristic, supervised by Gábor Recski
-
Multilingual hallucination detection for RAG applications
Verdha, supervised by Ádám Kovács
-
Large language model-based framework for open information extraction, triplet matching, and text comparison
Csakvari, supervised by Ádám Kovács
-
Open information extraction for fact-checking large language models
Osmanaj, supervised by Ádám Kovács
-
Rule learning for open information extraction
Sommer, supervised by Gábor Recski
-
Rule-based open information extraction from German legal domain
Iszak, supervised by Gábor Recski
2024
-
Advanced pattern matching in graph-based relation extraction: a methodical approach to improving XAI NLP systems
Piwonka, supervised by Gábor Recski
-
Evaluating LIME-based explanations of relation extraction models
Beham, supervised by Gábor Recski
Projects and collaborations
Research projects that lay the foundations of our technologies, including several collaborations with universities and the public sector. For the productised stack, see Technology.
VerbatimKG with WU SemSys
A collaboration with the WU SemSys group extending VerbatimRAG toward question answering over knowledge graphs and knowledge-graph population from trusted text sources. The project is funded by the Vienna Business Agency.
How to cite
BibTeX for individual papers lives on each paper page. A generic software-citation entry for the practice as a whole is below.
@software{krlabs-2026,
author = {{KR Labs}},
title = {KR Labs: open-source libraries for verifiable AI},
year = {2026},
url = {https://krlabs.eu/}
}Hear when we publish or ship
Occasional notes when we release a paper, model card, repository, project update, or collaboration. A few times a year, written close to the work.
Updates from the lab, sent only when there is something useful to say.
Read it, run it, cite it.
The code, the models, and the papers are open. The systems we deploy use the same architectures, and we support teams in defending them under audit.