The words that run AI-era engineering.

An AI agent is an LLM-powered system that pursues a goal autonomously — planning steps, calling tools, and iterating on results until the task is done.

AI red-teaming is the systematic practice of adversarially testing a model before deployment to find unsafe or policy-violating behavior and fix it first.

The Brier score measures probabilistic prediction accuracy: the mean squared gap between stated confidence and the outcome. 0 is perfect; lower is better.

Calibration is how well stated confidence matches reality — a calibrated person who says "90% sure" is right about 90% of the time.

Chain-of-thought prompting has a model write out intermediate reasoning steps before its final answer, improving accuracy on multi-step problems.

Chunking is splitting a document into smaller passages before embedding, so retrieval returns focused evidence instead of an entire file.

Citation and attribution links each claim in an AI answer to the source passage that actually supports it, so claims can be verified, not just trusted.

A code grader is a deterministic eval scorer — exact match, regex, schema check, or test pass/fail — that grades AI output without a human or LLM.

Context engineering is deciding what information enters a model's limited context window, and in what order, not just how instructions are worded.

A context window is the maximum number of tokens a model can read and generate in one request — the hard ceiling on what it can attend to at once.

Cosine similarity measures how alike two vectors' directions are — their dot product divided by the product of their magnitudes — from -1 to 1.

Cost per token is the price an API charges per token of input and output, usually asymmetric — output tokens cost several times more than input.

COSTAR is a prompt-writing checklist — Context, Objective, Style, Tone, Audience, Response format — that turns vague requests into complete, reliable prompts.

An embedding is a list of numbers (a vector) that represents the meaning of text, so that semantically similar texts end up numerically close together.

An eval gate is a CI check that blocks a deploy unless eval scores meet a defined threshold — a quality gate for prompts and models, not just code.

Few-shot prompting shows a model a handful of input-output examples in the prompt so it infers the pattern before answering a new case.

Function calling lets a model emit a structured request to run a specific function with specific arguments, which the application then executes.

A golden dataset is a curated set of inputs with human-verified expected outputs, used as ground truth to score an AI system in evals.

Grounding is how well an AI answer is supported by retrieved source material, rather than the model's own memory — the property RAG aims to guarantee.

A hallucination is when an AI model states false information with confidence — fabricating facts, citations, or code that looks plausible but is wrong.

HNSW is a graph-based approximate nearest-neighbor index that finds near-matches in roughly logarithmic time by searching multiple layers of shortcuts.

Human-in-the-loop is a design pattern where a person reviews or approves an AI system’s output at a defined checkpoint before it takes effect.

Hybrid search combines keyword search (like BM25) with vector search in one query, then merges the results — catching both exact matches and semantic matches.

In-context learning is a model's ability to infer and perform a new task purely from patterns shown in the prompt, without updating its weights.

An incident postmortem is a blameless written analysis of an outage — timeline, root causes, impact, action items — to prevent recurrence, not assign fault.

A jailbreak is a prompting technique that tries to get a model to bypass its safety training and produce output it was trained to refuse.

JSON mode constrains a model's output to be syntactically valid JSON — parseable, but not guaranteed to match any particular schema or fields.

The KV cache stores a transformer’s key and value vectors for every token already processed, so generating each new token skips recomputing the whole sequence.

Latency percentiles like p50, p95, and p99 show the response time a given share of requests fall at or under, exposing slow outliers a mean averages away.

Least-privilege tooling scopes every credential an AI agent can use to the minimum access it needs, so a mistake or attack causes bounded damage.

LLM evals are systematic tests that measure an AI system's quality — scoring outputs against defined criteria so changes are verified with data, not vibes.

LLM-as-judge uses a language model to grade another model's outputs against a rubric, scaling like code but validated against human labels.

Mastery learning is an approach where students must demonstrate mastery of each topic — typically 90%+ — before advancing; time varies, learning does not.

MRR is the average, across many queries, of 1 divided by the rank of each query's first relevant result — rewarding an early correct hit.

MCP is an open protocol standardizing how AI apps connect to tools and data, so each integration is built once instead of once per app-tool pair.

Model drift is the degradation of an AI system’s real-world performance over time, from a changed underlying model or a shift in production data.

Multi-agent orchestration coordinates several specialized AI agents working together on a task, instead of one agent handling everything serially.

A planner-executor loop splits agent work into a planning phase that decomposes a goal into steps and an execution phase that carries each step out.

Precision@k is the fraction of the top k retrieved results that are actually relevant — capped at 1/k when only one relevant document exists.

Prompt caching stores the static part of a prompt once so requests sharing that prefix pay a fraction of normal cost instead of reprocessing it every time.

Prompt engineering is the practice of designing model inputs — instructions, context, examples, and output format — to get reliable results from an LLM.

Prompt injection is an attack where untrusted content a model reads contains hidden instructions crafted to hijack its behavior away from the user’s intent.

Prompt regression testing reruns a fixed eval suite against every prompt change, blocking deploys that score worse than the prior baseline.

Quantization stores a model’s weights in a lower-precision format, like 8-bit or 4-bit integers instead of 16-bit floats, to shrink memory and speed inference.

ReAct interleaves reasoning ("Thought"), actions ("Action"), and results ("Observation") in one loop, so a model plans and acts together.

Recall@k is the fraction of all relevant documents for a query that appear somewhere in the top k retrieved results.

Reciprocal rank fusion merges multiple ranked search result lists by scoring each item as 1/(k + rank), summed across lists — no score normalization needed.

Reranking rescores a shortlist of retrieved candidates with a more accurate model before generation, moving the truly relevant result to the top.

RAG is a technique where an AI system retrieves relevant documents first, then generates an answer grounded in them — cutting hallucination sharply.

Semantic search ranks results by embedding similarity to a query's meaning, so paraphrases match even when they share no keywords.

SM-2 is the classic spaced-repetition scheduling algorithm: after each self-graded review, it updates an ease factor and computes the next interval.

Spaced repetition is a learning technique that schedules reviews at increasing intervals, timed just before you forget — producing durable memory.

Structured outputs constrain a model's decoding to guarantee its response matches a given schema exactly, not just probably resemble it.

Subword tokenization splits text into frequent word pieces rather than whole words, so any input can be represented without unknown-word gaps.

A system prompt is the hidden instruction set a developer supplies before a conversation starts, setting a model's role, rules, and boundaries.

A T-shaped engineer combines broad working knowledge across many domains (the horizontal bar) with deep expertise in one specialization (the vertical stem).

Temperature rescales a model's output probabilities before picking the next token — low values are near-deterministic, high values more random.

A token is the smallest unit of text a language model reads and writes — usually a word fragment of roughly 3-4 characters in English.

Tool use (function calling) is when a model emits a structured call to a function, which the surrounding application executes and returns a result for.

Top-p sampling draws the next token from the smallest set of candidates whose cumulative probability passes a threshold p, cutting the unlikely tail.

A vector database is a database built to store embedding vectors and find the nearest ones to a query vector at scale, in milliseconds.

Zero-shot prompting asks a model to perform a task from instructions alone, with no worked examples, relying on its pretrained knowledge.