AI Engineering Playbook
System Design

System Design

Everything above, assembled into one system, under constraints that conflict.

Prerequisites

Effectively the whole site — this section assembles it. At minimum:

Why this section exists

Designing an AI system is not a distributed-systems problem with an LLM sprinkled on top, and it isn't a matter of knowing the components either. It's holding five things in tension — quality, latency, cost, freshness, and safety — while staying explicit about what you'd measure.

Two failure modes dominate, and they're opposites:

  • Naming without structure. Pinecone, LangGraph, and a reranker get chosen before a single requirement is stated. Impressive vocabulary, no engineering.
  • Structure without commitment. Endless "it depends" and no decision. A design has to choose, defend the choice, and name what would change it.

How to reason about it gives the framework that avoids both: requirements → data → pipeline → serving → evaluation → operations. It's the thing that keeps an open-ended problem from stalling at the first step.

Key insight

Do the arithmetic explicitly. 10M documents × 4 chunks × 1536 dims × 4 bytes ≈ 245 GB of raw vectors — which doesn't fit one node's RAM, which forces sharding or quantization. Numbers are what turn a plausible-sounding architecture into a defensible one, and they usually decide the design for you before any preference does.

What's in here

PageThe question it answers
How to reason about itThe repeatable framework — read this before any of the worked designs
Design: Enterprise RAGRAG over 10M documents for 50k users — the canonical question
Design: Document Q&A at ScaleHigh-volume document QA with freshness and access control
Design: Semantic Code SearchSearch over a large codebase — why code isn't prose, and what that changes
Design: Customer Support AgentAn agent that resolves tickets, with tools, escalation, and guardrails
Design: LLM GatewayGoverned multi-model access: quotas, audit, routing — the enterprise middleware question

Where this connects

  • Rapid Fire — the compressed recall layer over everything here.
  • Every earlier section — when a design feels thin at some step, that step's page is the repair.

On this page