A lab notebook that cites its sources

The problem

Quantum hardware researchers produce two kinds of knowledge: measurements, recorded with QCoDeS, and reports, written in Lattice Notes, an open-source Microsoft markdown notebook. The two lived apart. Finding out what a colleague had already tried, and which data it was based on, usually meant a chain of messages between researchers.

Our team of five interns had full freedom to improve the measurement experience. My part was making that knowledge findable.

What I built

A retrieval system that connects datasets and reports:

• Datasets: when one is uploaded, a processing job extracts its metadata: protocol, device and instruments.
• Reports: when one is pushed to GitHub, it is chunked by section and paragraph and indexed for semantic search.
• Knowledge graph: everything is linked together, from report chunks and reports to datasets, protocols and instruments, so any claim can be traced back to the data behind it. Embeddings live in Azure AI Search; the graph lives in Azure Postgres.

Researchers don’t open a new app to use it. An MCP server exposes four tools to the agents they already use in VS Code, while they write reports or run experiments: semantic search, graph exploration, evidence search and grounded citation.

Key decisions

No evidence, no answer. An answer is only returned when evidence search succeeds, and every answer cites the report it came from. Researchers can’t afford made-up claims about lab data, so a cited answer or no answer at all was the only acceptable behaviour.

A graph on top of the embeddings. Semantic search alone finds similar text. The graph is what lets a researcher go from a paragraph in a report to the exact dataset and instruments behind it.

Built with the people using it. I had weekly feedback with my manager and an engineer in Copenhagen, and interviewed a different researcher every week to check it solved a real problem.

Results

A question that used to take a chain of messages between researchers took about a minute. I demoed it twice to the whole Quantum Delft office, including leadership.

It did not reach production use. As interns we only had a basic NDA, without clearance for protected quantum data, so it could not be connected to the real research archive.

The team

Alongside my work, my fellow interns built a self-learning agentic kernel, chunked cloud storage for datasets, and reproducible plots.

What I learned

Software is the bridge between cutting-edge research and the people doing it. In a lab, a tool that is sometimes wrong is worse than no tool at all, so trust has to be designed in from the start.