A grounded help-center chatbot case study — heading-aware chunking, a worked vector-index sizing calculation, a retrieval-plus-re-ranking pipeline, a citation-and-decline generation prompt, and an evaluation framework that scores retrieval and answer quality separately.
The project brief: a SaaS company's help center has 500+ published articles, and support wants a chatbot that answers customer questions grounded in those actual docs — citing its source so an answer can be verified, and cleanly saying “I don't know, here's how to reach a human” when the docs don't cover a question, instead of confidently guessing. That refusal behavior is the real engineering challenge here: a chatbot that answers well when the docs clearly apply is not hard to build, but one that recognizes when they don't is what separates a support tool worth shipping from one that generates its own support tickets.
This module walks the full build the way a production AI engineer would: heading-aware document chunking with a worked chunk-size and overlap decision, an embedding-model choice paired with a worked vector-index sizing calculation for this exact catalog size, a two-pass retrieval-and-re-ranking pipeline and why skipping the re-rank step measurably hurts precision on ambiguous questions, a grounding-and-decline system prompt, and an evaluation approach that scores retrieval precision/recall separately from final-answer faithfulness and refusal accuracy. The full worked numbers, complete tech-stack manifest, and finished design reasoning are part of the unlocked module below.