Harvey AI Interview Questions

The most-reported Harvey question by a wide margin, and the one the phone screen most often draws. Three progressive parts: count word-level matches of each source phrase in a model output, wrap the matched spans in <yellow> tags with overlapping and adjacent matches merged into one block, then append citation indices sorted by how often each source appears across the whole text. Interviewers correct candidates who count citation frequency per-highlight instead of globally.

The system design round that shows up in nearly every Harvey onsite report, sometimes framed as a PDF-only “Data Room”. Listing, upload/download, sharing, and access control, with the follow-ups converging on two things candidates lose the round on: organization-level ACL for sharing with thousands of users, and the unusual probe one candidate quoted verbatim — how do you verify that each chunk uploaded to S3 is actually yours.

Harvey’s other system design, described in one report as 他们家的基本盘 — their bread and butter. The reports carry an unusually specific failure mode: candidates prepare LLM and RAG internals, then lose the round on the crawler and the distributed-systems layer. One candidate was strong in every other round and failed on exactly that; another says outright to focus on crawler scale-up.