About NIZO
NIZO, founded in 1948 by the Dutch dairy industry, is an independent contract research organization for food and health based in Ede, the Netherlands. More than 100 scientists and support staff work on proteins, fermentation, food safety, and process upscaling, including in the largest food grade pilot plant in Europe available for third-party use.
The challenge: eight decades of knowledge behind a search bar that finds nothing
Since 1948, NIZO has compiled a vast archive of research reports, now digitized on SharePoint. However, the standard search function proved inadequate. Due to missing keywords and inconsistent search terms, valuable knowledge remained hidden.
NIZO wanted a simple search entry point that handles free-text questions and ranks documents from most to least relevant, so that newly hired employees could reach earlier work as well. Beyond finding a specific report, NIZO also wanted a discovery function that shows how its research is related. Newer AI techniques are well suited to making those connections visible, for example through a graphical explorer of the archive. NIZO looked for a partner that could first demonstrate which retrieval technique suits an archive full of specialist terminology spanning eight decades, rather than building a system straight away.
The solution: two techniques compared, one built out
DataNorth AI built the solution in short sprints, together with a project group of five NIZO employees from research and business development. The strength of the approach was the decision to build and assess two retrieval techniques side by side before developing one of them further, in close consultation with Eric Hester, Bioinformatics Lead at NIZO.
The approach consisted of:
- Comparison of two techniques: Both GraphRAG and a rerank approach were built and evaluated. The rerank approach was selected for its substantially lower upfront cost and its better fit with what NIZO needed.
- Hybrid search engine: Our AI Engineers combined keyword search (BM25) with vector search and had the combined results reordered by a rerank model, using OCR for reports without a text layer.
- Visual exploration: Alongside the ranked result list, the purpose was to get discovery based on relatedness of work, so an interactive cluster diagram of the search space was added, with filters for document type, year, and research area among others.
The results: impact in figures
After two sprints, the solution was tested in daily practice by a project group of five NIZO employees:
Delivered in the proof of concept
- More than 75 years of research reports are searchable through a single entry point.
- Saves 2 hours of onboarding time per employee.
- Expected 20% faster research process
Conclusion
By first testing which retrieval technique fits its own archive, NIZO gained a working search solution in two sprints that opens up more than 75 years of research. The collaboration with DataNorth AI proves that a short proof of concept with real users gives direction faster than a long preliminary study.
