What to Expect in the Harvey System Design Interview

Harvey's system design round asks you to design systems like the ones behind its legal AI products. Reported question themes include document processing pipelines, search systems, collaboration features, high volume APIs, and permission systems. A permission system decides which user may see which document. The round rewards candidates who connect the design to legal users, where confidentiality and accuracy are strict requirements.

Harvey builds generative AI for law firms. Its products read large sets of legal documents and answer questions about them with citations. The design questions come from exactly that work.

The Question Types Harvey Asks

  • Document pipelines. Ingest thousands of files, extract text, split it into chunks, and index it. A chunk is a small piece of text sized for a model to read.
  • Search and retrieval. Combine keyword search with vector search. Vector search finds text by meaning, using numeric representations called embeddings.
  • Permissions and confidentiality. Law firms separate matters strictly. Access control lists (lists of who may access each item) come up often in this round.
  • Collaboration systems. Several lawyers edit and comment on the same document at once.
  • High volume APIs. Serve model-backed endpoints with rate limits, retries, and predictable latency.

Here is a high level path through a representative question: design a search system over a law firm's documents.

Step 1: Requirements. Ask about scale (documents, users, queries per second) and about freshness. Then state the legal constraints yourself: results must respect permissions, and answers must cite sources. Naming those constraints early is the strongest signal in this round.

Step 2: Ingestion. Files arrive in many formats. A pipeline extracts text, detects structure, splits it into chunks, and writes to two indexes. One index is keyword based, one is vector based. Make ingestion a queue-driven pipeline so a bad file cannot stop the flow. Store the original file too, because lawyers will ask to see the exact source page.

Step 3: Query path. A query runs against both indexes. Merge the results and rank them with a scoring model. Filter by permissions before ranking, not after. Filtering after ranking risks leaking a document's existence to the wrong user.

Step 4: Answering with citations. Pass the top chunks to a language model and require it to cite the chunks it used. Reject answers whose claims lack a supporting chunk. This is the main defense against hallucinations (false statements presented as fact).

Step 5: Evaluation and scale. Keep a test set of real questions with known answers and score every release against it. Track citation accuracy, not just relevance. For scale, cache frequent queries, shard the indexes, and isolate each firm's data completely. Sharding means splitting one index across many machines.

What Interviewers Grade

They grade constraint thinking first. Permissions and accuracy are not optional features in legal software. Candidates who raise them before being asked do well.

They grade trade-offs spoken aloud. Keyword versus vector search, latency versus quality, cost versus freshness. Name both sides and choose with a reason. A stated assumption you can defend beats a silent guess every time.

They grade communication. Follow a clear order: requirements, high level design, one deep area, risks, measurement. Ask the interviewer which area to expand. This round is one of four onsite rounds, so the loop context in What Is the Harvey Interview Process Like? (Round by Round) is worth reading first.

How to Prepare

TAGS
System Design Interview
CONTRIBUTOR
Arslan Ahmad
Arslan Ahmad
ex-FAANG engineering manager and author or Grokking series.
-

GET YOUR FREE

Coding Questions Catalog

Design Gurus Newsletter - Latest from our Blog
Boost your coding skills with our essential coding questions catalog.
Take a step towards a better tech career now!
Explore Answers
Who is the current CEO of Amazon?
Best practices for designing scalable backend services
Master the best practices for designing scalable backend services. Covers stateless design, connection pooling, database scaling, resilience patterns, and production-grade architecture.
What is Docusign interview process reddit?
What is your goal in life?
Explain DocuSign system design interview.
What type of system is Zoom?
Related Courses
New
Grokking the AI System Design Interview course cover
Grokking the AI System Design Interview
Learn to design AI systems the way interviewers expect: classic ML products, LLM and RAG architectures, and agentic systems, all through the lens of the system design interview.
4.6
(3,192 learners)
Discounted price for Your Region

$123

Grokking the Coding Interview: Patterns for Coding Questions course cover
Grokking the Coding Interview: Patterns for Coding Questions
The 24 essential patterns behind every coding interview question. Available in Java, Python, JavaScript, C++, C#, and Go. The most comprehensive coding interview course with 543 lessons. A smarter alternative to grinding LeetCode.
4.6
Discounted price for Your Region

$197

Grokking Modern AI Fundamentals course cover
Grokking Modern AI Fundamentals
Master the fundamentals of AI today to lead the tech revolution of tomorrow.
4.1
Discounted price for Your Region

$72

Design Gurus logo
One-Stop Portal For Tech Interviews.
Copyright © 2026 Design Gurus, LLC. All rights reserved.