How a custom RAG over 76 years of UN documents helps nearly 100,000 students prepare for Model UN conferences with primary sources.
Model Diplomat is a generative AI research tool purpose-built for Model United Nations (MUN). Founded in 2023 by CEO Georgina Songhurst, the platform helps nearly 100,000 students worldwide conduct rigorous, source-backed research as they prepare position papers, opening speeches, and committee strategy.
Kinematic Labs partnered with Georgina as a technical AI partner to architect a production-grade retrieval-augmented generation (RAG) system spanning more than two million United Nations documents from 1946 through 2022, turning seven decades of multilateral diplomacy into a queryable, citation-first knowledge base.
For thousands of students, Model UN is the first place they encounter the real work of diplomats: drafting resolutions, defending national positions, and negotiating across blocs. The materials that make MUN substantive include real UN speeches, working papers, voting records, and committee reports. They constitute some of the richest diplomatic primary sources in the world.
They are also some of the least accessible. Most of the UN's institutional record sits in decades of digitized PDFs of varying quality, with no consumer-grade research tool indexed against them.
Georgina Songhurst, a longtime MUN participant, saw the gap firsthand. Students were turning to general-purpose AI tools that confidently invented quotes, misattributed positions, and hallucinated votes. The shortcut was actively undermining the educational experience MUN was designed to deliver.
"Students deserved an AI grounded in the real diplomatic record, not a chatbot that guessed, but a research companion that pointed delegates back to what nations had actually said and done."
Georgina Songhurst, CEO, Model Diplomat
While the UN's institutional record is one of the richest civic data sets in the world, it is one of the least machine-readable. Records dating back to 1946 exist primarily as scanned PDFs, many predating modern OCR standards and spanning multiple languages, formats, and typographies.
Off-the-shelf AI models had no exposure to this corpus. When asked about specific votes, statements, or working papers, they produced answers that were fluent, plausible, and frequently wrong.
2M+ PDFs spanning 76 years of varying scan quality, multilingual content, and historical typography predating digital publishing.
The system needed to surface and cite specific documents, paragraphs, and votes with precision no general-purpose model could offer.
In an educational context, fabricated quotes and misattributed positions actively undermine the learning experience.
Infrastructure needed to scale gracefully to a worldwide audience of students preparing for conferences simultaneously.
Georgina engaged Kinematic Labs as her technical AI partner to architect and build the system. From the first working session, she set two non-negotiable product principles that shaped every engineering decision that followed:
Fidelity over fluency. The tool should refuse to answer rather than fabricate.
Citations are first-class. Every response must lead students back to the source.
These principles drove the choice of architecture, the design of the retrieval layer, and the way the user experience surfaces sources to students.
"I needed a partner who understood that for Model Diplomat, the technology had to serve the truth. Kinematic Labs delivered by adhering to two absolute non-negotiables: fidelity over fluency, and citations as a first-class feature, ensuring students are always grounded in primary sources."
Georgina Songhurst, CEO, Model Diplomat
Model Diplomat runs on a production-grade RAG stack engineered for fidelity at scale. Each layer was selected to reinforce the founding principles.
76 years of UN PDFs processed through a high-precision OCR pipeline tuned for variable-quality scans, multilingual documents, and historical typography.
The cleaned corpus of over two million documents sits in Amazon S3, structured and indexed for downstream retrieval.
Pinecone provides low-latency semantic search across the full corpus, with chunking strategies tuned for the formal cadence of diplomatic language.
Cohere's reranker selects the most relevant passages from initial retrieval candidates, dramatically improving precision before content reaches the LLM.
The OpenAI API generates responses constrained tightly to retrieved context, with citation surfaces built into every output.
When this system shipped in 2023, large-scale production RAG was rare. Model Diplomat was one of the most ambitious civic-knowledge RAG deployments of its era.
Model Diplomat has grown to nearly 100,000 students preparing position papers, opening speeches, and committee strategy with grounded, source-backed research.
Students engage more deeply with primary sources. Conference organizers see better-prepared delegates. The platform has become a fixture of MUN preparation in classrooms, clubs, and universities around the world.
"Seeing nearly 100,000 students move beyond generic, hallucinated answers to engage with the actual diplomatic record has been transformative. It proves that when you prioritize rigour over speed, you're not just building a product, you're elevating the entire standard of civic education."
Georgina Songhurst, CEO, Model Diplomat
Under Georgina's leadership, Model Diplomat continues to expand, deepening corpus coverage, refining the retrieval layer, and exploring how grounded AI can support civic education well beyond Model UN.