Summer 2025/26
Knowledge agent
A retrieval-augmented agent that answers questions over the club's own document corpus instead of guessing.
The club accumulates documents — prospectuses, planning docs, run sheets, handover notes — and nobody can find anything in them. This team built an agent that answers questions against that corpus directly.
The core of it is retrieval-augmented generation: rather than relying on what a model already knows, the system finds the relevant passages in the club's own documents first and answers from those. When it doesn't have the material, it's supposed to say so rather than invent an answer.
What the team had to work out
- Chunking documents so retrieval returns useful passages, not fragments
- Getting the model to defer to retrieved context over its own priors
- Keeping answers traceable back to a source document
Other builds
Build the next one
Applications for each projects intake go out to members first. No experience assumed — there's a beginner track for exactly that reason.