I Built an AI That Knows My Watch Collection (RAG Demo)

A general AI knows plenty about Seiko watches, but nothing about my collection: when I bought each watch, when each one is due for service, which straps fit which watch, or my own rules for wearing them in water. In this demo I built a Retrieval-Augmented Generation (RAG) assistant that answers those questions from a custom knowledge base instead of guessing. How it works Retrieval – My Collector's Handbook is split into sections, each labeled with an ID. A search that matches both meaning and keywords finds the sections most relevant to each question. Augmentation – The code automatically builds a prompt that combines the retrieved sections, strict "answer only from this context" instructions and today's date. Generation – Claude Haiku answers using only that context and cites the handbook sections it used (e.g. [C3]). In the demo – Which watch is overdue for service (date math across two sections) – Why a vintage diver rated for 150 m still stays out of the pool (house rule vs. manufacturer rating) – Which straps fit which watch – A question the handbook doesn't cover, which the bot declines to answer instead of making something up – The same question with and without RAG, side by side Chapters 0:00 The problem 0:30 The open-book idea 0:50 Retrieval 1:30 Augmentation 2:00 Answers with citations 3:15 With vs. without RAG Built with Python, Google Colab, sentence-transformers, scikit-learn and the Anthropic API (Claude Haiku 4.5). Created for an AI course project. The collection records in the handbook are sample data; the Seiko specs and warranty terms come from Seiko's official documentation.