RM Labs

RM Labs is a UX concept for a document Q&A assistant: drag in whatever files you already have — contracts, policies, a renewals spreadsheet — and ask questions in plain English instead of hunting through pages by hand. The interesting constraint was trust, not retrieval: every sentence in an answer carries a numbered citation, and hovering or clicking it cross-highlights the exact retrieved passage and page it came from, so nothing in the answer is left unverifiable.

Technical highlights

• 01

  Built a fully scripted, dependency-free interaction sequence (Promise-chained async steps with a cancellation token) that drives a three-pane workspace — sources, conversation, and retrieved passages — through file upload, indexing, question typing, retrieval, and token-by-token answer streaming.

• 02

  Designed bidirectional citation linking: inline citation pills in the streamed answer and their source passage cards share hover/focus/click state, so a reader can jump from a claim straight to the sentence that grounds it and back.

• 03

  Implemented an accessible file-scope control (keyboard-operable checkbox list with ARIA roles) that visibly changes what the assistant is allowed to search, making the retrieval boundary a first-class, inspectable part of the UI rather than a hidden backend detail.

• 04

  Respected `prefers-reduced-motion` end-to-end (typing, streaming, count-up, and transition timings all collapse under a shared time-scaling helper) and kept the whole demo replayable and interruptible via a single monotonically incrementing run token.