Skip to main content

AI From Zero to Builder

Glossary

Canonical, learner-first definitions for recurring terms used across the curriculum, labs, and project work.

Use this page when a technical word interrupts the lesson. Definitions are intentionally short: enough to recover the idea and continue learning, while the Lessons provide the deeper context.

53 of 53 terms

A

admission controlProduction
A rule that decides whether new work may start now, wait in a queue, or be rejected based on available capacity and limits.
agent harnessAgents
Application code around an agent that manages state, tools, permissions, retries, traces, budgets, and other execution rules.
artifactAgents
A result produced by a task, such as a file, report, structured record, or other output that can be tracked separately from messages.

B

backoffReliability
Waiting longer before another retry so repeated failures do not cause a rapid retry loop.
backpressureProduction
Slowing or limiting new work when downstream parts of a system are already near capacity.
blast radiusSecurity
The amount of data, users, systems, or actions that one failure or compromised component could affect.

C

canaryProduction
A small early release that sends limited traffic to a candidate version before a wider rollout.
capabilityProtocols
A specific action, resource, interface, or skill that a system says it can provide.
cardinalityObservability
The number of distinct values or label combinations in a dataset or metric; very high metric cardinality can be expensive to store and query.
cosine similarityRetrieval
A similarity score that compares vector direction after accounting for vector length.

D

data exfiltrationSecurity
Protected information leaving a system and reaching a destination that is not authorized to receive it.
data minimizationPrivacy
Collecting, sending, and keeping only the data needed for a declared purpose.
degraded modeProduction
A reduced version of a service that keeps the most important work running when full capacity or functionality is unavailable.
delegationAgents
Giving a bounded subtask to another agent or worker while keeping limits on its tools, budget, authority, and expected result.
dot productMath & retrieval
A number produced by multiplying matching vector components and adding the results; it can reflect both alignment and vector magnitude.
durable stateReliability
State stored so it can still be recovered after a process, worker, or machine stops.

E

embeddingRetrieval
A vector representation that places items in a numeric space so useful relationships can be compared mathematically.
evictionProduction
Removing stored or cached data to free limited memory for other work.

H

headroomProduction
Capacity deliberately left unused during normal operation so the system can absorb bursts, retries, or failures.

I

idempotency keyReliability
A stable identifier that lets repeated attempts be recognized as the same intended operation instead of new operations.
inferenceModels
Running a trained model on input to produce predictions, scores, embeddings, or generated output.

J

jitterReliability
A small variation added to retry timing so many workers do not all retry at exactly the same moment.

K

KV cacheProduction
Stored attention key and value state from earlier tokens that lets later generation steps reuse work instead of recomputing all prior attention state.

L

latencyProduction
How long a request or operation takes from a chosen start point to a chosen completion point.
least privilegeSecurity
Giving a user, agent, tool, or process only the permissions and data access needed for its current job.
load sheddingProduction
Rejecting or dropping some lower-priority work when accepting everything would overload the system.
localityProduction
Keeping data close to the compute that uses it so transfers cost less time and bandwidth.

O

observabilityObservability
Collecting and connecting enough traces, metrics, and logs to explain how a system behaved without guessing.
operation IDReliability
An identifier for one intended operation that can connect retries, completion evidence, and audit records.

P

p95Evaluation
The 95th percentile: a value that 95% of measured observations are at or below.
preemptionProduction
Pausing or removing active work so limited resources can be given to other work.
prefillProduction
The generation phase that processes the input tokens already present before one-token-at-a-time decoding continues.
principalSecurity
The trusted user, service, or other identity on whose behalf an action would occur.
provenanceEvidence
Information about where a piece of data, result, or claim came from and how it reached the current system.

Q

quantizationModels
Representing model numbers with fewer bits or a more compact numeric format to reduce storage or computation costs, with possible quality and compatibility tradeoffs.

R

reconciliationReliability
Checking durable or external evidence to determine what actually happened before deciding whether to repeat an uncertain operation.
rerankingRetrieval
Scoring a smaller set of retrieved candidates more carefully to improve their final order.
residual riskSafety
Risk that still remains after the planned controls have been applied.
rollbackProduction
Routing or deploying future work back to an earlier known version after a candidate release fails its gates.

S

sandboxSecurity
An execution environment that limits what files, network destinations, devices, or resources a process can reach.
schemaInterfaces
A machine-readable description of the fields, types, and structure that data is expected to follow.
side effectReliability
A change outside the current computation, such as charging money, sending a message, deleting a file, or updating a record.
sliceEvaluation
A group of evaluation cases inspected separately because its behavior may matter on its own, such as one task type or risk category.
spanObservability
One timed operation inside a trace, such as queue wait, model inference, or a tool call.
statelessProtocols
Designed so one request carries the information needed at that boundary instead of depending on hidden per-session protocol state.

T

tenantSecurity
One customer, organization, or isolated account boundary in a system that serves multiple groups.
thrashingProduction
Repeatedly moving, evicting, or recomputing data so aggressively that the overhead reduces useful progress.
threat modelSecurity
A structured description of what must be protected, which actors and boundaries exist, how failure or attack could happen, and where controls should stop it.
throughputProduction
How much useful work a system completes per unit of time, such as requests or tokens per second.
traceObservability
A connected record of the operations that one request or task passed through.
trust boundarySecurity
A place where data or requests cross between components that should not automatically trust one another.

V

vector normMath & retrieval
A measure of vector length; in cosine similarity it is used to normalize magnitude.
version preconditionReliability
A rule that accepts an update only if the stored version still matches the version the writer previously read.