Sources — public references

The ideas on this site are not new; they are assembled. Below are the public documents the chapters lean on, grouped by theme, with one line on why each matters here. Every link is public and free to read.

Serving and scaling at consumer scale

• Optimizing AI Inference at Character.AI↗ — a consumer chat product describing the caching and batching decisions that make serving tens of thousands of turns per second affordable; the concrete backdrop for prefix caching and "the bill is tokens".
• Efficient Memory Management for LLM Serving with PagedAttention↗ — the paper behind modern inference servers; explains why memory, not compute, bounds how many concurrent users one model replica can hold.
• Prompt caching (provider documentation)↗ — a worked description of prefix caching from a model provider: what is cacheable, how prompts should be ordered, and the price difference.
• Building effective agents↗ — a plain-spoken taxonomy of agent patterns; the source for the "model decides, code executes" framing used in the tools chapter.
• Effective context engineering for AI agents↗ — why a curated, budgeted context beats a stuffed one; the argument behind the context pack.

Retrieval and memory

• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks↗ — the original RAG paper: retrieve first, then generate, so answers are grounded in fetched material rather than model weights.
• Retrieval-Augmented Generation for Large Language Models: A Survey↗ — a map of the RAG design space (query rewriting, re-ranking, fusion) that the nine-stage turn draws from.
• Generative Agents: Interactive Simulacra of Human Behavior↗ — the memory stream scored by recency × importance × relevance; the direct ancestor of temporal decay in personal memory.
• Lost in the Middle: How Language Models Use Long Contexts↗ — evidence that models under-use material buried mid-prompt; the reason a bigger context window is not the same as a better answer.

Control planes and autoscaling

• Kubernetes: Horizontal Pod Autoscaling↗ — the reference description of the HPA: target metric, desired-replica arithmetic, cool-downs, scaling in both directions.
• Kubernetes: Controllers↗ — the desired-state / actual-state control loop stated in its simplest form; the pattern the reconcile loop chapter teaches from first principles.
• Cloud Run: About instance autoscaling↗ — a managed autoscaler's actual rules (concurrency, CPU, min/max instances, scale to zero); the production counterpart of the hand-built loop.
• Cloud Run: Rollouts, rollbacks, and traffic migration↗ — revision-level traffic splitting, which is how a canary is expressed on a managed platform.
• Google SRE Book: Service Level Objectives↗ — where SLOs and error budgets come from, and why they belong in the quality chapter rather than a contract.

Tools and the Model Context Protocol

• Model Context Protocol — Specification↗ — the normative text: initialize, tools/list, tools/call, and the error model; the basis for the N×M → N+M argument in tools.

Evaluation

• Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena↗ — the paper that made "a model grades the model" a respectable eval method, and measured where it agrees with people; the basis for the two-score eval gate.
• Your AI Product Needs Evals↗ — a practitioner's account of building golden sets from real failures and gating releases on them; the everyday version of the golden set discipline.

Vector search at scale

• Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs↗ — the HNSW index most vector stores use; explains the recall-vs-speed trade that makes vector search feasible at millions of vectors.
• pgvector↗ — a vector index inside a relational store; the reason per-user memory can live next to the profile it belongs to, without a second system.
• Vertex AI Vector Search overview↗ — what a dedicated, managed vector service looks like when an unscoped corpus outgrows an in-database index.