Grounded AI for teams that cannot afford made-up answers.
We build AI systems whose answers point back to source text, for teams that have to defend their output.
Live now Evidence-grounded QA over the ACL AnthologyWhat we do
Research-led delivery.
KR Labs is a small team of NLP researchers and engineers. We publish the methods, maintain the open-source libraries, run the audits, and build the systems clients deploy. There is no handoff between theory and implementation.
Verifiable AI systems.
Our stack is built around evidence: VerbatimRAG for source-grounded answers, LettuceDetect for unsupported spans, and RuleChef for explicit domain logic. The tools are open-source, model-agnostic, and designed to be inspected.
Built for regulated work.
We work with teams in finance, legal, healthcare, public sector, and research, where AI output has to survive review. The goal is not a better demo. It is a system whose answers, evidence, and controls can be defended under audit.
Three ways in
The same practice, three entry points. Read the one that fits where you are.
Practice
How KR Labs engages commercially. Audit-led, deployment-oriented, three weeks to a written report. Then we decide together what to build next.
[technology]Technology
Three product lines that constitute a deployable stack: VerbatimRAG, LettuceDetect, RuleChef. Fully open-source and model-agnostic.
[research]Research & Insights
Papers, whitepapers, and posts from the same people who run the audits and ship the code. Read what we publish, then read what we deploy.
The technology family
Three open-source architectures that constitute a deployable stack. MIT and Apache 2.0 licenses across the family.
[verbatim]
For answers that have to survive audit. Every generated span carries a pointer back to a verbatim source passage a reader can verify. Model-agnostic.
[lettucedetect]
Flags the texts that your LLM may have made up. Token-level, encoder-based, 30 times smaller than prompt-based detectors. F1 79.22 on RAGTruth.
[rulechef]
When system logic has to be explicit and inspectable. Learns regex, Python, and spaCy patterns from your examples; runs locally with no LLM at inference.
Live now
Current KR Labs platforms, demos, and integrations you can use today.
Verbatim for evidence-grounded QA.
A hosted platform built around our Verbatim product, starting with trustworthy question answering over the ACL Anthology. Every answer is assembled from exact source spans with citations.
- Corpus
- 114,484 papers
- Answers
- exact source spans
- Access
- MCP-ready
A working example: the evidence trail
A defensible answer is not just generated. It is assembled from source spans, marked with citations, and traceable back to the exact passages that support each claim. The highlighted path shows how an answer connects back to its evidence.
A query is grounded against source documents. Span extraction selects verbatim evidence, template assembly and LLM completion produce a cited answer, and each citation points back to the source span that supports the claim.
Selected research and insights
Three starting points from the research archive. The full archive lives at Research & Insights.
Defensible by design, not by promise.
Start with an audit of the AI system you already have. Three weeks. Senior practitioners. A written report with findings, evidence, and architecture recommendations. Then we decide together what to build next.