ChatGPT Enterprise (RAG ML System Design)

Design ChatGPT Enterprise: enterprise users upload internal data; the system gives them a customized ChatGPT based on that data. The Challenge Build an intelligent chatbot system that uses Retrieval-Augmented Generation (RAG) to answer user queries — a business tool in the style of Glean that searches company data and synthesizes clear, sourced answers using an LLM. Main Goals More than basic RAG: full chatbot system with multi-user support, permissions, and data safety — not a tutorial-level prototype. Business requirements: multi-tenancy, access control (who can see what), and citation tracking. Answer quality: responses must be grounded in retrieved data, with hallucination prevention and source attribution. Latency: fast retrieval without sacrificing answer quality. Requirements Be ready to discuss each of these technical areas in depth: Embeddings: how to produce text embeddings; which models (contrastive learning + hard-negative sampling); trade-offs between dense and sparse representations. Vector database: selection criteria (Pinecone, Weaviate, Chroma, etc.); ANN index choices (HNSW); tenant-scoped index vs. shared index + filter. Chunking strategy: how to split large documents for retrieval; chunk size trade-offs; metadata extraction per chunk. Retrieval strategy: semantic search vs. hybrid (dense + sparse) retrieval; when to use each; handling factual vs. open-ended queries; handling vague or ambiguous questions. Reranking: pointwise vs. pairwise vs. listwise reranking; model options (cross-encoder, LambdaMART); when to add a reranking stage. Prompt engineering: how to construct prompts so the LLM uses retrieved context correctly and does not fabricate. Citations: how to track and surface source attribution end-to-end. Caching: embedding cache; semantic cache for popular queries; latency impact. Evaluation metrics: retrieval recall/precision; end-to-end response quality; hallucination detection. Conversation history: how to maintain and inject session context across turns. Security and access control: ensuring users see only permitted documents; multi-tenancy isolation. System Design Key structural decisions to articulate: Pipeline separation: keep ingestion (chunking, embedding, indexing) cleanly separate from retrieval and generation; each scales independently. Component boundaries: how ingestion service, retrieval service, and generation service communicate; API shape between them. Three-stage retrieval: candidate generation → ranking (e.g., LambdaMART or cross-encoder reranker) → LLM reranking/synthesis. Multi-tenancy isolation: tenant-scoped index or per-tenant namespace with access control enforcement at retrieval time. Chat history storage: database schema for session logs; how to bound context window usage. Scaling: batch processing for document ingestion; horizontal scaling of retrieval; handling millions of documents. Examples Retrieval architecture: dense + sparse hybrid retrieval 3-stage: candidate generation → LambdaMART ranking → LLM reranking Chunking strategy, metadata extraction, multi-tenancy isolation Evaluation: retrieval recall/precision, end-to-end response quality Drilled point: pointwise vs listwise reranking tradeoff. A common flame-out pattern: propose pointwise, get pressed on weaknesses, drift into listwise joint-optimization and inter-document dependencies without crisp scoping — end-of-loop feedback comes back as "some gaps." Have the listwise comparison ready before the question lands. Preparation Full RAG: embedding model (contrastive learning + sampling), ANN index (HNSW), hybrid retrieval, reranking Pointwise / pairwise / listwise reranking comparison Multi-tenancy isolation: tenant-scoped index + access control Latency reduction: embedding cache, semantic query cache, batch ingestion Hallucination mitigation: grounding constraints in prompt, fallback when no relevant docs retrieved Evaluation: RAGAS or equivalent; offline recall/precision + online quality signals Reference reading Public architecture guides worth skimming before the loop (build your own design from requirements — don't memorize these verbatim): Designing High-Performing RAG Systems Microsoft Azure — RAG Solution Design and Evaluation Guide Galileo AI — Mastering RAG: Enterprise RAG Architecture AWS — What is Retrieval-Augmented Generation? Notes Common mistakes to avoid Treating it as a tutorial: the interviewer expects enterprise-level thinking — authentication, authorization, session tracking, observability. Single-turn focus: the system must handle multi-turn conversation; forgetting conversation history management is a common gap. No evaluation plan: failing to specify how you'd measure retrieval quality and answer correctness signals weak ownership. Ignoring scale: no plan for millions of documents, high-concurrency retrieval, or ingestion throughput. Ignoring safety: omitting data privacy, access control, and tenant isolation.