Back to course overview
Module 8Capstone project 12 min

App design

Your own corpus, your own users: choosing the document set, the design document, and the corpus audit that decides everything downstream.

HarborDocs was the rehearsal. Your capstone is a document Q&A application over a corpus you actually have — team documentation, a product knowledge base, policy manuals, research notes, a course's readings. Same architecture, real stakes: this is the project you'll demo, and for many alumni the one that goes to work on Monday.

Choosing the corpus (the decision that matters most)

  • 20–500 documents you have the right to process — enough for retrieval to matter, small enough to audit by hand. Mind access control: if the docs carry permissions, scope the capstone to a single-permission slice.
  • Questions actually get asked of it. A corpus nobody queries makes an eval set of fiction. Best sign: you can list ten real questions from memory because people keep asking you.
  • Some messiness required. All-clean-markdown teaches nothing; all-scanned-PDFs teaches only pain. One or two format challenges is the sweet spot.

The corpus audit (before any code)

  1. 1Inventory: formats, sizes, structure quality, duplication, staleness. Which extraction paths do you need? Which documents are traps (tables, images-of-text, tracked changes)?
  2. 2Collect 15 real questions users have asked of this material — from tickets, chats, or memory. These seed the golden set and define the product.
  3. 3Decide the metadata schema from those questions (the two-question test per field), and the chunking strategy per document type.
  4. 4Name the risks: PII? Access control? Stale-document hazards? Write the mitigation next to each — your Foundations data-class training applies to corpora too.

The design document (deliverable 1)

Two pages: corpus audit summary; users and their top questions; metadata + chunking plan; retrieval plan (hybrid? reranker? — with the why, or the honest 'baseline first, upgrades if metrics demand'); trust layer configuration (FLOOR policy, confidence routing, citation UX); success criteria as numbers (recall@5 ≥ X, verification ≥ Y, abstention correctness ≥ Z on unanswerables). Every choice traceable to the audit or the questions — 'because the tutorial did it' appears nowhere.

Design review, the cheap way

Run your design doc past your assistant with the hostile-reviewer prompt pattern you know: 'find the three weakest assumptions'. Then past one human who knows the corpus. Ten minutes each; both will find something the other can't.