AnyoneLearnAI
LevelEasyStart

Reference

Glossary A–Z

Look up any AI term — search or jump by letter. Each entry links to a lesson when one exists.

190 terms

A

  • A/B testingSplitting traffic between **two variants** (prompt, model, retrieval) and comparing metrics to pick
  • Adapter (LoRA)A **small trainable layer** bolted onto a frozen base model — cheaper than full fine-tuning.
  • Adversarial promptA user or document crafted to **trick an AI** into breaking rules — ignoring policies, calling forbi
  • AgentAn AI loop that plans, uses tools, observes, and continues toward a goal.
  • Agent memoryWhat an **agent loop remembers** across steps — message history, tool results, and optional long-ter
  • AGINews headlines, research debates, and safety discussions about long-term AI capability — not when de
  • AISoftware that learns patterns from data to make predictions, classifications, or generated content —
  • AI copyrightAI copyright concerns how copyright law applies to AI training data, generated outputs, and human au
  • AI literacyBaseline skills to use, question, and improve with AI tools safely.
  • AI monitoringAI monitoring tracks a deployed AI system's quality, reliability, safety, and operational performanc
  • AI regulationLaws, standards, and agency rules that govern **how AI systems are built, deployed, labeled, and aud
  • AlignmentShaping AI systems to **follow human intent and values** — helpful, honest, harmless — not just pred
  • Anthropic APICloud API for **Claude models** — messages API with long context, system prompts, tool use, and visi
  • API keyA secret string that **authenticates your app** to an AI provider's API — like a password for your p
  • API rate limitProvider cap on requests per minute — batch jobs help large embed/index builds stay under limits.
  • Audit trailImmutable **log of who changed what** — prompts, models, indexes — for compliance and debugging.

B

  • Baseline metricA **reference number from stable production** — last week’s p95 latency or citation rate — used to d
  • Batch inferenceRunning many model calls in one scheduled batch instead of one sync request at a time.
  • Batch jobAn async job that processes many API requests offline — poll until complete.
  • Batching (inference)Running **multiple model requests together** in one forward pass to improve GPU throughput — often a
  • BenchmarkA **standardized test suite** used to compare models on tasks like math, coding, safety, or retrieva
  • BiasSystematic **unfair skew** in AI outputs — often reflecting biased training data or design choices t

C

  • Canary deploymentReleasing a change to **a small slice of traffic first**, watching metrics, then promoting or rollin
  • Chain-of-thought promptingChain-of-thought prompting asks a model to work through intermediate reasoning steps before giving a
  • ChatbotAn app that **accepts user messages and returns AI replies** — often with memory, tools, or retrieve
  • Chunk overlapRepeated text between adjacent chunks so sentences split across boundaries still retrieve.
  • ChunkingSplitting long documents into **smaller pieces** before embedding or retrieval so RAG systems can fi
  • Citation (in AI answers)Pointing to **where a claim came from** — a doc title, URL, chunk ID, or page — so readers can verif
  • ClassificationPredicting a **discrete label** — spam/not, yes/no, cat/dog — from inputs or scores.
  • ClusteringClustering is an unsupervised learning method that groups data points by similarity.
  • CompletionThe model's **generated continuation** of a prompt — the next tokens it predicts, whether a single p
  • Computer visionAI that **interprets images and video** — detect objects, read text (OCR), segment regions, estimate
  • Confusion matrixA table of **true vs predicted labels** — shows false positives and false negatives, not just overal
  • ConsentPermission before using someone's likeness, voice, or data with AI.
  • Context windowThe maximum amount of text (measured in **tokens**) a model can read and remember **in one conversat
  • Conversation memoryTechniques to keep multi-turn chats useful within **token limits** — trim, summarize, or store facts
  • Cosine similarityA score from **-1 to 1** measuring how aligned two vectors are — common for comparing embeddings.
  • Cost allocationTagging **token spend by team, feature, or customer** so you know who consumed what.
  • Cost per tokenHow providers **price API usage** — separate rates for input tokens (your prompt) and output tokens
  • Cross-encoderA model that **scores query–document pairs together** — slower than bi-encoders but sharper for re-r
  • Cross-validationSplitting data into **multiple train/test folds** so you estimate performance more reliably than a s

D

  • Data driftWhen **live inputs change** from what you trained or indexed on — new slang, products, or user behav
  • Data governanceRules for who can access, store, and train on organizational data.
  • Data pipelineAutomated flow from raw data → clean tables → features → model or index.
  • Dataset splitDividing examples into **train / validation / test** so you measure generalization, not memorization
  • Dead letter queueStorage for **failed automation jobs** after retries exhaust — so nothing silently disappears.
  • Decision treeA decision tree predicts an outcome by applying a sequence of feature-based splits.
  • Deep learningDeep learning uses multilayer neural networks to learn representations and patterns from data.
  • DeepfakeAI-generated or altered **video, audio, or images** that make a real person appear to say or do some
  • Diffusion modelA generative model that **starts from noise and iteratively denoises** into an image (or other signa
  • DistillationTraining a **smaller "student" model** to mimic a larger "teacher" model's outputs — keeping much of

E

  • EmbeddingNumbers that represent the meaning of text so computers can compare ideas.
  • Embedding batchGrouping many texts into **one embeddings API job** (or batch endpoint) instead of one call per row.
  • Embedding indexStored vectors for chunks — built once, queried at retrieval time.
  • Embedding modelA model that turns text into **fixed-size vectors** for similarity search — not for chat completions
  • Eval gateA **CI or deploy checkpoint** that blocks release when golden tasks score below threshold.
  • Eval setA fixed list of **input → expected-quality** tasks used to compare prompts, models, or releases.
  • Experiment trackingLogging **prompt, model, data, and metric** for each training or eval run so you can compare and rep
  • Exponential backoffRetry strategy: **wait longer after each failure** (1s, 2s, 4s…) to avoid hammering a rate-limited A

F

  • FairnessFair treatment across groups in AI outcomes — not identical outputs for everyone.
  • FaithfulnessHow closely a model output **matches the provided context** — no extra claims beyond what sources su
  • Feature engineeringChoosing and transforming **input columns** so a model can learn — scaling, encoding categories, or
  • Few-shot promptingGiving the model **a few input→output examples** in the prompt so it copies the pattern on new input
  • Fine-tuning
  • Function callingWhen an LLM **returns a structured request** to run a function (API, database, calculator) instead o

G

  • Generative AIAI that **creates new content** — text, images, audio, video — from a prompt or input, rather than o
  • GitGit is a distributed version control system for tracking and combining changes to files.
  • GitHubGitHub is a platform for hosting Git repositories and collaborating on software.
  • Golden taskA **fixed test input** with known quality criteria — the backbone of eval gates and regression tests
  • GPUGraphics Processing Unit — hardware that **parallelizes matrix math**, making neural network inferen
  • Gradient descentAn optimization loop that **nudges model weights** in the direction that reduces loss — the core “le
  • GroundednessWhether an answer **sticks to retrieved or cited sources** instead of inventing facts.
  • GroundingTying an answer to **specific source text or data** so the model is less likely to invent facts.
  • GuardrailsRules, filters, and checks that **limit what an AI system can say or do** — content policies, tool a

H

  • HallucinationWhen an AI states something **false but sounds confident** — invented citations, wrong dates, fake p
  • Hugging FaceAn open platform for **models, datasets, and inference** — hub hosting, Transformers library, and sp
  • Human in the loopHumans approve high-stakes AI outputs before they take effect.
  • Hybrid searchCombining **keyword search** (exact matches, BM25) with **semantic/vector search** so you catch both

I

  • IdempotencyProcessing the **same event twice** without duplicate side effects — critical for webhook retries.
  • Incident responseStructured steps when **production AI fails** — contain harm, rollback, investigate, communicate.
  • Inference
  • Inference endpointHTTP (or gRPC) URL where a deployed model accepts inputs and returns predictions.
  • Input filterServer-side checks on **user content before it reaches the model** — pattern blocks, length limits,

J

  • JailbreakA prompt or trick that **bypasses an AI's safety rules** — getting it to ignore policies, reveal sec
  • JSON modeAPI setting that nudges the model to reply with parseable JSON instead of prose.
  • JSON schemaA contract describing allowed JSON keys and types — models and validators use it for machine-readabl

K

  • KV cacheExplaining why long contexts cost more and why caching prefixes helps latency.

L

  • LangChainA popular **Python/JS framework** for chaining LLM calls — prompts, retrievers, tools, agents, and m
  • LatencyHow long users wait for an AI response — from sending a prompt to seeing the first token (TTFT) or t
  • LLMA Large Language Model predicts the next piece of text from patterns it learned.
  • LLM tracingRecording **prompts, tool calls, latencies, and outputs** per request so you can debug production AI
  • Local inferenceServing a model on your machine or VPC instead of a public cloud API.
  • LoRACustom tone, domain style, or task format when you have example data but not a huge GPU budget.
  • Loss functionA loss function measures how far a model's predictions are from the desired outcomes during training

M

  • Machine LearningA branch of AI where programs **improve from examples** instead of following only hand-written rules
  • matplotlibPython’s common **plotting library** — scatter plots, histograms, and line charts to see patterns be
  • Max tokensAn API setting that caps **how long the model's reply** can be — measured in tokens, not words.
  • MCPA standard way to plug tools and data into AI apps — like USB for AI capabilities.
  • Metadata filter (vector search)Restrict similarity search to chunks matching **tags** — tenant, date, doc type — before ranking vec
  • Model cardA standard doc summarizing a model's limits, data, and intended use.
  • Model routerLogic that **picks which model** handles each request — fast/cheap for drafts, capable/expensive for
  • Model routingSending each request to the right model size or provider for cost and quality.
  • Model servingModel serving makes a trained model available to applications for inference, usually through an API.
  • Model weightsThe **learned numbers** inside a neural network that encode patterns from training — billions of par
  • Multi-agent systemA multi-agent system coordinates multiple autonomous agents that interact to pursue shared or separa
  • Multimodal

N

  • Narrow AIAI that excels at **one task or domain** — spam filtering, chess, speech-to-text — without general h
  • Neural networkLayers of weighted connections that learn patterns from data.
  • NLPSearch, translation, summarization, sentiment analysis, chatbots, and entity extraction from documen

O

  • ObservabilityMeasuring **logs, metrics, and traces** from live AI systems so you detect drift, cost spikes, and q
  • OllamaA popular tool to **download and run open-weight LLMs locally** with a simple CLI and API.
  • On-callThe **rotation that responds** to production alerts outside business hours.
  • On-device AIModels running on your hardware — data stays local, useful for privacy and offline.
  • Open weightsModel **parameters published for download** so you can run or fine-tune on your own hardware — licen
  • Open-source modelPublicly released model weights you can run locally or fine-tune — check license terms.
  • OpenAI APICloud HTTP API for **ChatGPT-class models** — chat completions, embeddings, images, audio, and assis
  • Orchestrator agentAn agent that **plans, delegates, and merges** work from specialist worker agents instead of running
  • Output validationChecking model responses against **schema, regex, or policy rules** before showing or storing them.
  • OverfittingWhen a model **memorizes training examples** but fails on new data — great accuracy in the lab, poor

P

  • P95 latencyThe response time below which **95% of requests** finish — catches slow tail better than average alo
  • pandasA Python library for **tables of data** — load CSVs, filter rows, compute stats, and prep features f
  • PEFTCustom style, format, or domain phrasing when full fine-tune is too expensive.
  • PostmortemA **blameless write-up** after an incident — timeline, root cause, and follow-up actions.
  • Pre-trainingThe expensive first stage where an LLM reads massive text to learn general language patterns — befor
  • Precision@kFraction of **top-k retrieved chunks that are relevant** — measures noise in the context window.
  • PrivacyKeeping personal data out of models and logs unless policy allows.
  • Production AI architectureProduction AI architecture is the system design that connects models, data, applications, infrastruc
  • PromptThe instructions and context you give an AI model — your question, task, examples, and constraints.
  • Prompt cachingReusing **cached prefix tokens** (system prompt, long docs) so repeat calls pay less for the same co
  • Prompt engineeringThe practice of **designing instructions, examples, and constraints** so models produce reliable, us
  • Prompt injection

Q

  • Quality gateAn automated **check that blocks release** when quality metrics fail — golden queries, citation chec
  • QuantizationCompressing model weights to **fewer bits** (e.g. 16-bit → 8-bit or 4-bit) so models run faster and

R

  • RAGRetrieve relevant documents first, then generate an answer from them.
  • Random forestA random forest combines many randomized decision trees to produce a more robust prediction.
  • Rate limitA cap on **how many API requests or tokens** you can use per minute — providers enforce it to keep s
  • Re-rankingA second pass that **re-scores retrieved chunks** with a smarter (often slower) model so the best pa
  • Reasoning modelAn LLM configuration that spends **extra compute on stepwise thinking** before answering — better on
  • Recall@kWhether the **correct chunk appears in the top k** retrieved results — retrieval quality before gene
  • Red teamingDeliberately **attacking your own AI system** to find failures — unsafe outputs, leaks, bad tool cal
  • Reinforcement learningTraining by **trial and reward** — an agent takes actions, gets feedback, and learns policies that m
  • Responsible AIBuilding and using AI with safety, fairness, and accountability in mind.
  • RetrievalThe step that **finds relevant documents or chunks** before the model generates an answer — the "R"
  • Risk tierClassification of AI use by potential harm — stricter rules for higher tiers.
  • RLHF
  • RollbackReverting a **live AI deployment** to the last known-good model, prompt, or retrieval index when qua
  • Root cause analysisFinding the **underlying failure** — bad index, prompt, model, or infra — not just the symptom.
  • RunbookA **step-by-step playbook** for common incidents — who to page, what to rollback, where logs live.

S

  • Schema validationChecking parsed JSON against required keys/types before your app acts on it.
  • Semantic cacheReturning a **stored answer** when a new question is semantically near a past query — skip LLM call
  • Semantic searchSearch by **meaning**, not just exact keywords — powered by embeddings and similarity.
  • Service recoveryRestoring **normal operation** after an incident — rollback, hotfix, or traffic shift back to health
  • Similarity searchFinding items whose **meaning** is close to a query — usually by comparing embedding vectors with di
  • SLO (service level objective)A **target for reliability or quality** — e.g. p95 latency under 2s or citation rate above 85% — tha
  • Span (tracing)One **timed unit of work** in a trace — e.g. embed query, retrieve chunks, call chat completion.
  • Speech recognitionSpeech recognition converts spoken language in audio into text.
  • SSE (Server-Sent Events)An HTTP pattern where the **server pushes chunks** to the browser over one long response — common fo
  • Streaming (LLM responses)Sending model output **token-by-token** (or chunk-by-chunk) over HTTP instead of waiting for the ful
  • Structured outputForcing the model to reply in a **machine-readable shape** — JSON, XML, or a fixed schema — instead
  • Supervised learningMachine learning from **labeled examples** — each input comes with the correct answer the model shou
  • Sync APIRequest/response API for interactive latency — user waits for one result.
  • Synthetic dataBootstrapping eval sets or formatting examples when real data is scarce.
  • Synthetic mediaContent **created or heavily modified by AI** — images, video, audio, text — especially when it coul
  • System messageThe API role that sets **global instructions** for the assistant — tone, rules, and scope for the wh
  • System promptHidden instructions that set the AI's **role, rules, and tone** for a whole session — before the use

T

  • TemperatureA dial that controls how **random** vs **predictable** an LLM's word choices are. Low = safer and re
  • Text-to-imageAI that **generates pictures from text prompts** — diffusion and related models map words to pixels.
  • ThresholdA **cutoff score** that turns a continuous model output into a yes/no (or class) decision.
  • TokenA small chunk of text the model reads and writes.
  • Token budgetA **cap on tokens** per request, user, or day — prevents runaway spend from long contexts or agent l
  • TokenizerThe piece of an LLM stack that **splits text into tokens** — the units models read, price, and count
  • Tool / function callingLet an AI app run real actions (weather, calendar, API) instead of only chatting.
  • Tool allowlistAn explicit list of **tool names the model may call** — everything else is rejected at runtime.
  • Top-kKeeping only the k highest-scoring retrieval hits before re-rank or generation.
  • Top-p (nucleus sampling)A sampling setting that limits the model to the **smallest set of likely next tokens** whose combine
  • Traffic splitSending only a **fraction of requests** to a new version while the rest stay on stable — core of can
  • TransformerA **transformer** is the architecture behind modern LLMs. Its **attention** mechanism weighs which e
  • TransparencyClear disclosure when content or decisions involve AI.

U

  • Unsupervised learningFinding **patterns in data without labels** — grouping similar items, detecting anomalies, or reduci
  • User messageThe API role for **what the human (or app) asks** in this turn — the main task input.

V

  • VectorA list of numbers that can be treated like a point in space.
  • Vector databaseA database optimized to **store embeddings and run similarity search** fast — Pinecone, pgvector, Ch
  • Vector indexData structure (HNSW, IVF, etc.) for fast nearest-neighbor search over embeddings.
  • Virtual environmentA virtual environment isolates a Python project's packages and versions from other projects and the
  • Vision encoderThe part of a **multimodal model** that turns images into tokens the LLM can reason over.

W

  • WebhookAn **HTTP callback** when an event happens — new row, form submit, payment — triggering your automat
  • Workflow automationChaining **triggers, AI steps, and app actions** (email, CRM, Sheets) with human checkpoints.
  • Workflow orchestrationCoordinating **multi-step automations** — triggers, LLM calls, branches, retries — as one durable fl
  • Workflow triggerThe **event that starts an automation** — form submit, new email, schedule, webhook.

Z

  • Zero-shot promptingAsking the model to do a task **with instructions only** — no example pairs in the prompt.
← Reference hub