Design a RAG system for vehicle documents
Company: Sonatus
Role: Software Engineer
Category: System Design
Difficulty: hard
Interview Round: Technical Screen
Design a Retrieval-Augmented Generation (RAG) architecture to power a Q&A chatbot over **vehicle documentation**.
### Document types & characteristics
- Sources: engine specification manuals, warranty documents, repair documents, vehicle datasets.
- Formats: PDFs (1 to 1,000 pages), CSVs (potentially large), HTML, Markdown.
- Workload: **read-heavy**.
- Updates: documents are **static** (assume no edits after publishing, but new document versions may be added).
### Core requirements
1. Users ask natural-language questions (e.g., “What torque spec for …?”, “Is this covered under warranty?”).
2. System retrieves relevant passages/snippets and generates an answer grounded in the documents.
3. Provide citations (doc + page/section/row reference where possible).
4. Handle both unstructured text (PDF manuals) and structured/semi-structured data (CSV datasets).
### Non-functional requirements (discuss and propose targets)
- Latency and throughput goals for chat requests.
- Quality: relevance, correctness, citation accuracy.
- Security: document access control if documents differ by vehicle model, region, or user entitlements.
- Observability: tracing, retrieval diagnostics.
- Scalability: growing corpus size.
### Deliverables
- High-level architecture (ingestion/indexing + query-time flow).
- Data model for documents/chunks/metadata.
- Vector database choice/strategy and retrieval approach.
- How to process PDFs/CSVs for chunking and citations.
- Key tradeoffs and failure modes.
Overview: This question evaluates a candidate's ability to design a Retrieval-Augmented Generation system, testing competencies in information retrieval, vector search, document ingestion and chunking, metadata modeling, citation handling, security, observability, and scalability within the system design and NLP/information-retrieval domain.