RAG knowledge layer for an AI product

KiwiStart is an AI product that helps migrants and students settle in New Zealand. Its chat only works if the model retrieves the right facts before it answers, and that is a data problem before it is a model problem.

I worked on the retrieval layer: structuring the knowledge base so answers could be grounded, and tuning the embeddings and vector search that pick what the model sees for a given question. The goal throughout was retrieval quality, getting the right context in front of the model and keeping wrong or off-topic content out.

This is the part of AI most analytics people never touch. It is also where my data-engineering habits transfer directly: clean sources, clear structure, and checks on quality. A retrieval layer built like a proper data pipeline is the difference between an AI that sounds confident and one that is actually right.

What I worked on

• Structuring the knowledge base so content could be retrieved and cited reliably.
• Embeddings and vector similarity search to match questions to the right source material.
• Improving retrieval relevance to cut down wrong or off-topic answers.

Scope note: written to reflect real contribution without overstating. Exact responsibilities can be described more precisely on request.