The Knowledge-Context Protocol for Chemistry

Large language models are increasingly applied to diverse scientific tasks, but their performance degrades on knowledge outside their training data, and they rarely provide verifiable proof for their assertions. While retrieval-augmented generation supplies missing knowledge, it often returns disconnected text and offers little guarantee that a retrieved fact was actually used. We present the Knowledge-Context Protocol (KCP), a Model Context Protocol-based format that packages an ontology as a callable cartridge paired with semantic tools. This framework allows for assertions to be tracked and verified at inference time. We further pair KCP with automated natural language benchmark construction from the information in the ontologies. We evaluate nine open-weight and frontier models on 440 chemistry questions on acid base conjugation and reaction classification. When evaluated on QA accuracy, access to the KCP cartridge shows improvement over the tested baselines. Tool use analysis shows that correct answers are predominantly those reaching the required ontological entities. While the KCP is only applied to chemistry in this case, we expect the approach to transfer to other domains with formal ontologies.