What is an AI glossary?

An AI glossary is a reference list defining the vocabulary used across artificial intelligence and automation — model architectures, prompting techniques, agent patterns, and the platform-specific words that surround them. This one covers 483 terms in plain English, with each definition linked to the tools or templates that actually use it.

What AI terms should I learn first?

Start with the eight that everything else is built on: large language model, token, context window, prompt, hallucination, RAG, AI agent, and workflow automation. Almost every other term in this glossary is a refinement of one of those, so learning them first makes the rest read as variations rather than as new concepts.

What is the difference between AI, machine learning and deep learning?

They nest. Artificial intelligence is the broad goal of software doing things that normally need human judgement. Machine learning is the dominant way of achieving it — learning patterns from examples instead of following written rules. Deep learning is a kind of machine learning that uses many-layered neural networks, and it is what made modern language and vision models work.

How often is this AI glossary updated?

Continuously. The agent, MCP and skills vocabulary in particular is only months old and still settling, so terms are added as they stabilise rather than on a fixed schedule. Definitions are written in-house rather than aggregated, so each one can be corrected when the usage shifts.

Is this AI terminology guide free?

Yes. All 483 terms are on one page with no signup, no account and no paywall, and the eight most-searched terms have their own full explanations.

AI Glossary

483 AI and automation terms, defined in plain English. A complete AI terminology guide covering large language models, AI agents, MCP servers, RAG, prompt engineering, machine learning and the automation platforms they run on — each term linked to the tools and templates that put it to work.

Definitions are written in-house rather than aggregated, so each one says what the term means in practice and what it costs you to get wrong — not a paraphrase of the same sentence every other AI dictionary carries.

Free and complete on one page — no signup, no gate. Last updated August 2026.

Start here — the terms worth a full explanation

What is an n8n workflow?What is an MCP server?What is an AI agent skill?What is workflow automation?What is an AI agent?What is RAG (retrieval-augmented generation)?What is prompt engineering?What is a webhook?

A

Action: A step a workflow performs after being triggered: create a record, send a message, transform data, call an API. Workflows chain actions to complete a process.
API (Application Programming Interface): A defined way for software to talk to other software. Automation platforms are essentially friendly wrappers around thousands of APIs.
AI agent: An LLM given tools, memory, and a goal, allowed to plan and execute multi-step work with limited supervision. Full explanation →
AI benchmark: A standardized test (MMLU, HumanEval, GPQA…) for comparing model capability. No single benchmark tells the whole story; look at profiles across many. Live AI benchmarks →
AI wrapper: A product whose core is a prompt plus a third-party model API. Thin wrappers compete on UX and distribution, not model quality.
Automation ROI: Hours saved × loaded hourly cost, minus platform and build costs. Calculate before building — some automations are not worth their maintenance. ROI calculator →
AI agent skill: A packaged, reusable instruction set — usually a Markdown file — that teaches an agent how to perform one specific task, loaded on demand rather than kept in context permanently. Full explanation →
A/B test: A controlled experiment that randomly exposes comparable users or requests to different variants and measures outcome differences. Statistical significance and guardrail metrics are essential.
A/B testing: Splitting traffic between variants to measure which performs better. The only reliable way to attribute a change to an outcome.
Access token: A credential presented to an API to authorize requests for a limited scope and period. Access tokens must be protected like passwords.
Adapter: A small set of trainable parameters added to a base model to specialize it without updating every original weight. Adapters make customization cheaper and easier to swap.
Agent card: A machine-readable description of an agent's identity, capabilities, endpoints, authentication requirements, and supported interaction modes used for discovery and coordination. Explore the AI agents directory →
Agent discovery: The process of finding an agent and understanding what it can do before delegating work. Discovery may use registries, agent cards, configuration, or trusted directories. Explore the AI agents directory →
Agent handoff: Transfer of a task, context, and responsibility from one agent to another. A good handoff includes the goal, completed work, evidence, constraints, and expected output. Explore the AI agents directory →
Agent interoperability: The ability of independently built agents to discover, communicate, exchange tasks, and collaborate through shared protocols and data formats. Explore the AI agents directory →
Agent loop: The recurring cycle of observing state, deciding or planning, invoking a tool, receiving the result, and continuing until completion, escalation, or a stopping limit. Explore the AI agents directory →
Agent memory: The general capability of carrying information beyond a single prompt. Almost always a retrieval system, not a property of the model. Explore the AI agents directory →
Agent orchestration: Coordinating multiple agents or tools — routing work, sequencing, and merging results. Explore the AI agents directory →
Agent swarm: A large collection of agents coordinating through shared rules, messages, or an environment. Swarms can provide parallelism but are difficult to control, evaluate, and make cost-efficient. Explore the AI agents directory →
Agent2Agent Protocol (A2A): An open standard for communication and interoperability between independent AI agents built with different frameworks or by different vendors.
Agentic: Describing a system that decides its own next step toward a goal, rather than following a fixed script. Explore the AI agents directory →
Agentic loop: The repeating cycle of think, act, observe that an agent runs until it finishes or hits a limit. The limit matters — without one, failure looks like an infinite bill. Explore the AI agents directory →
Agentic workflow: A workflow that delegates some sequencing or decision-making to an AI agent while retaining explicit process boundaries. It is more flexible than a fixed workflow and more controllable than open-ended autonomy. Explore the AI agents directory →
AI alignment: Work on making model behaviour match human intent, including under instructions nobody anticipated.
AI automation: The use of AI models inside automated processes to interpret information, make bounded decisions, generate content, or choose actions. Unlike fixed-rule automation, AI automation can handle ambiguity but requires evaluation, guardrails, and fallbacks. Browse 27,000+ ready-made workflows →
AI gateway: A control layer between applications and model providers that can centralize authentication, routing, logging, rate limits, caching, policy enforcement, and cost tracking.
AI governance: The policies, roles, decision rights, controls, and evidence used to ensure AI systems are accountable, lawful, secure, and aligned with organizational goals.
AI model: A mathematical system learned from data that maps inputs to predictions, classifications, generated content, or actions. Model behavior depends on architecture, training, prompts, tools, and runtime settings.
AI risk management: The continuous process of identifying, measuring, prioritizing, treating, monitoring, and communicating AI risks throughout the system lifecycle.
AI slop: Low-effort mass-generated content published for volume. The reason search engines increasingly discount thin generated pages.
AI supply-chain risk: Risk introduced through third-party models, datasets, libraries, adapters, prompts, tools, or hosting providers. Provenance, integrity checks, version pinning, and vendor assessment reduce exposure.
Alert: A notification triggered when a metric, event, or evaluation crosses a threshold. Good alerts are actionable and avoid overwhelming operators with noise. Browse 27,000+ ready-made workflows →
Algorithm: A fixed procedure for turning input into output. In AI the algorithm is the training method; the thing you ship is the model it produced.
Answer relevance: How directly and completely an answer addresses the user's question without unnecessary or unrelated material. A relevant answer can still be factually wrong, so relevance is not sufficient alone.
Anthropomorphism: Attributing human feelings or intent to a model. Encouraged by chat interfaces and a common source of misplaced trust.
API endpoint: A specific network address representing an API resource or operation, such as POST /orders. Each endpoint defines accepted inputs, authentication, behavior, and outputs.
API key: A secret value used to identify and authorize an application calling an API. Keys should be scoped where possible, stored in a secret manager, rotated, and never placed in public content.
Application programming interface (API): A defined interface that lets software exchange data or invoke capabilities programmatically. APIs are generally more reliable and maintainable for automation than imitating clicks in a user interface. Browse 27,000+ ready-made workflows →
Approval gate: A checkpoint that prevents a workflow from continuing until an authorized person or policy grants approval. Gates should show the proposed action, evidence, risk, and consequences clearly. Browse 27,000+ ready-made workflows →
Artificial general intelligence (AGI): A hypothetical system matching human performance across essentially all cognitive tasks. No agreed definition or test exists, which is why claims about it are unfalsifiable.
Artificial intelligence (AI): Software that performs tasks normally requiring human judgement — recognising images, writing text, making predictions. In practice today it almost always means machine learning, not hand-written rules.
At-least-once delivery: A guarantee that a message will be delivered one or more times. It reduces loss risk but requires consumers to handle duplicates safely.
At-most-once delivery: A guarantee that a message is delivered no more than once. It avoids duplicates but may lose messages if a failure occurs before processing completes.
Attended automation: Automation initiated or supervised by a person on their device to assist with a task. It keeps the user in control and is common in customer service or back-office work. Browse 27,000+ ready-made workflows →
Attention mechanism: The operation letting a model weigh which other tokens matter when interpreting each one. The core idea that made transformers work.
Audit log: A tamper-resistant record of who or what performed an action, when it occurred, what resource was affected, and the outcome. Audit logs support accountability and investigation.
Authentication: The process of proving the identity of a user, application, or service. Authentication answers who is calling; authorization determines what that identity may do.
Authorization scope: A named permission limiting what an access token can do, such as read contacts or send email. Workflows should request the minimum scopes needed.
Automation return on investment (ROI): The economic value created by automation compared with build, platform, operation, maintenance, error, and review costs. A simple starting formula is benefits minus costs, divided by costs. Browse 27,000+ ready-made workflows →
AutoML: Automating model and hyperparameter selection. Reduces expertise needed, not the need to understand the data.
Autonomous agent: An agent that pursues a goal over many steps without per-step human approval. The risk scales with how many irreversible actions it can take. Explore the AI agents directory →
Autonomy level: The degree to which an AI system may decide and act without approval. Autonomy should increase only as task reversibility, observability, reliability, and permission controls improve.
Autoscaling: Automatically increasing or decreasing compute or worker capacity based on load or queue depth. Autoscaling must account for startup time, rate limits, and sudden demand spikes.
Availability: The proportion of time a system is usable. High availability uses redundancy, health checks, failover, capacity planning, and recovery procedures.

B

Backpressure: A mechanism that slows or rejects incoming work when downstream capacity is saturated. Without backpressure, queues and memory can grow until the system becomes unstable.
Backpropagation: The algorithm that assigns credit for an error back through every layer so weights can be adjusted. The engine of essentially all neural network training.
Base model: A pre-trained model before instruction tuning or alignment. Better at raw completion, worse at following directions.
Batch API: Submitting many requests for asynchronous processing at a lower rate, in exchange for waiting.
Batch inference: Submitting many model requests for asynchronous processing together, often at lower cost. It is appropriate when results are not required immediately.
Batching: Combining multiple items into one request or job. Batching can reduce overhead and cost, but large batches increase latency, memory use, and the impact of partial failures.
Beam search: Keeping several candidate continuations and choosing the best overall sequence. Common in translation, rare in chat.
Benchmark: A standard task set used to compare models. Useful for ranking, unreliable as a predictor of performance on your workload. Compare 298 models →
Bias: Systematic skew in data or model behaviour that disadvantages some group. Comes from the data far more often than from the algorithm.
Black box AI: A model whose internal reasoning cannot be inspected. A compliance problem wherever decisions must be explained.
BLEU: An overlap-based score for machine translation. Cheap and reproducible; poorly correlated with human judgement on modern systems.
Branch or conditional: Logic that sends execution down different paths based on a rule, such as order value, confidence score, or customer status. Branches turn a linear workflow into a decision tree. Browse 27,000+ ready-made workflows →
Business process automation (BPA): Automation of an end-to-end business process spanning people, rules, and systems, such as employee onboarding or invoice approval. BPA is broader than automating a single task. Browse 27,000+ ready-made workflows →

C

Context window: How much text (measured in tokens) a model can consider at once. Bigger windows fit more documents but cost more per call.
Claude Projects / rules: Persistent instructions and knowledge attached to Claude conversations so the assistant follows your conventions every time. Claude rules directory →
Cursor rules: Project-level instruction files that steer the Cursor AI code editor — coding style, framework conventions, forbidden patterns. Cursor rules directory →
ComfyUI workflow: A node-graph pipeline for Stable Diffusion image generation, shareable as JSON — the image-generation equivalent of an automation workflow.
Custom workflow: An automation built to your exact spec instead of adapted from a template — the right call when the process is core to your business. Request a custom build →
Caching: Storing a result so repeated requests can be served faster and more cheaply. Cache keys, freshness, invalidation, permissions, and sensitive-data isolation determine whether caching is safe.
Canary deployment: Releasing a change to a small portion of traffic first and expanding only if metrics remain healthy. Canary releases reduce the blast radius of prompt, model, or workflow changes. Browse the prompt library →
Capability negotiation: The handshake where client and server agree which protocol features each supports.
Chain-of-thought prompting: Prompting that encourages intermediate reasoning before an answer. In production, request concise rationales or verifiable intermediate results rather than depending on hidden reasoning as proof of correctness. Browse the prompt library →
Chatbot: A conversational interface. Modern ones wrap a language model; older ones matched keywords against scripted replies. Explore AI tools →
Checkpoint: A persisted snapshot of progress and state from which execution can safely resume. Checkpoints are especially useful for long-running workflows, agents, and large batch jobs.
Choreography: Decentralized coordination in which services react to events from one another without a central controller. It reduces central coupling but can make end-to-end behavior harder to trace.
Chunk overlap: Repeating some content between adjacent chunks so important context is not split at boundaries. Too much overlap increases index size and produces redundant retrieval results.
Chunking: Dividing documents into retrievable segments. Effective chunks preserve enough context to be meaningful without mixing unrelated ideas or exceeding the model's useful context.
Circuit breaker: A reliability pattern that temporarily blocks calls to a failing dependency after a threshold is reached. It periodically tests recovery instead of allowing every request to fail slowly.
Citation: A reference linking a model claim to supporting evidence. Useful citation systems verify that the source exists, is relevant, and actually supports the associated claim.
Classification: Predicting which category something belongs to. Evaluated with precision, recall and a confusion matrix — never accuracy alone on imbalanced data.
Cloud-hosted automation: Using automation infrastructure operated by a provider. It reduces operational work but introduces vendor dependency, data-location considerations, usage pricing, and service limits. Browse 27,000+ ready-made workflows →
Clustering: Grouping similar records without labels. The groups are only as meaningful as the distance measure you chose.
Cold start: Extra delay when an idle function, container, worker, or model must initialize before handling a request. Provisioned capacity or warm pools can reduce cold-start latency.
Compensating transaction: An action that logically reverses a completed step when a later step fails, such as issuing a refund after inventory reservation fails. It is not always a literal database rollback.
Computer vision: Getting software to interpret images and video — classification, detection, segmentation, OCR.
Concept drift: When the real-world relationship a model learned changes, so accuracy decays even though nothing in the code changed.
Concurrency: The number of operations in progress during the same period. Higher concurrency improves throughput until databases, APIs, or workers reach capacity.
Confusion matrix: The table of predicted versus actual classes. Shows which specific mistakes a model makes, which a single accuracy score hides.
Connector or integration: A prebuilt component that handles authentication and common operations for a service. Connectors accelerate setup but may expose fewer features than the service's complete API. Browse 27,000+ ready-made workflows →
Constitutional AI: Training a model against an explicit written set of principles, using the model itself to critique and revise its outputs.
Context caching: Reusing previously processed model context across requests to reduce repeated input processing, latency, or cost. Provider rules determine what is cacheable and for how long.
Context compression: Reducing context size while preserving task-relevant information through summarization, extraction, filtering, or structured state. Compression lowers cost but may remove important details.
Context engineering: Designing what information reaches a model, when it is included, how it is structured, and what is omitted. It includes prompts, retrieval, memory, tools, summaries, and state management.
Context injection: Programmatically adding instructions, retrieved documents, state, or tool results to a model request. Injected external content should be clearly delimited and treated as untrusted data.
Corpus: A body of text used for training or evaluation. Usually described by size and source, both of which determine what the model will and will not know.
Cosine similarity: A measure of the angle between two vectors, commonly used to compare embedding direction while reducing the influence of magnitude. Explore AI tools →
Cost guardrail: A rule that prevents or escalates work when expected or accumulated spending exceeds a threshold. Examples include model downgrades, tool-call caps, or human approval for expensive runs.
Cost per successful task: Total model, tool, infrastructure, and review cost divided by the number of tasks completed successfully. This is more meaningful than token cost alone when comparing systems.
Cost per task: Total spend divided by tasks completed successfully. The honest metric — cost per API call flatters a system that retries a lot.
Cron expression: A compact notation defining recurring schedules through time fields such as minute, hour, day, month, and weekday. Time zone and daylight-saving behavior must be specified explicitly.
Cross-validation: Rotating which slice of data is held out for testing, then averaging. Gives a more honest estimate than a single split when data is scarce.
Cutoff date: The point after which a model has no training knowledge. Anything later must be supplied in the prompt or retrieved. Browse the prompt library →

D

Data mapping: Matching fields between two systems (e.g., form "full_name" → CRM "contact.name"). Most of the real work in building integrations.
Data augmentation: Expanding a dataset with transformed copies — crops, paraphrases, noise — to improve robustness without new labelling.
Data exfiltration: Unauthorized extraction of sensitive information from a system. In AI automation, exfiltration may occur through model output, tool calls, logs, retrieval, connectors, or injected instructions. Browse 27,000+ ready-made workflows →
Data freshness: How current the information available to a workflow or model is. Freshness depends on source update frequency, ingestion delay, caching, and indexing speed. Browse 27,000+ ready-made workflows →
Data ingestion: Moving data from source systems into a processing, storage, analytics, or retrieval environment. Ingestion may be batch, streaming, scheduled, or event-driven.
Data minimization: Collecting, sending, retaining, and exposing only the data required for a defined purpose. Minimization reduces privacy risk, prompt size, breach impact, and compliance burden. Browse the prompt library →
Data mining: Extracting patterns from large datasets. Predates modern AI and is often statistical rather than learned.
Data poisoning: Manipulating training, fine-tuning, evaluation, or retrieval data to introduce bias, backdoors, misinformation, or unsafe behavior into an AI system. Explore AI tools →
Data residency: The geographic location where data is stored or processed. Residency requirements can affect provider, region, architecture, backup, and cross-border transfer choices.
Data transformation: Changing data's structure, format, or values so another step can use it. Examples include splitting names, converting dates, aggregating rows, or reshaping JSON.
Dead-letter queue (DLQ): Storage for messages or jobs that repeatedly fail processing. A DLQ preserves failed work for inspection, correction, and controlled replay instead of silently discarding it.
Decision tree: A model of nested yes/no splits. Individually weak but highly interpretable, and the building block of forests and boosting.
Decoder: The half that generates output one token at a time. Chat models are decoder-only.
Deduplication: Detecting and suppressing repeated events, records, or requests using identifiers, hashes, or time windows. Deduplication is crucial when delivery systems may send the same event more than once.
Deep learning: Machine learning using neural networks with many layers. It is what made image recognition and language models work, at the cost of needing far more data and compute. Explore AI tools →
Deepfake: Synthetic audio or video convincingly depicting a real person. The main driver behind provenance and watermarking work.
DELETE: An HTTP method used to remove a resource. Production automation should confirm identity, authorization, reversibility, and audit requirements before deletion. Browse 27,000+ ready-made workflows →
Deserialization: Converting transmitted or stored data back into program objects. Untrusted data must be validated during deserialization to prevent errors or security issues.
Deterministic automation: Automation in which the same inputs and rules produce the same expected result. It is best for calculations, routing, validation, and other processes that require predictable behavior. Browse 27,000+ ready-made workflows →
Differential privacy: Adding calibrated noise so individual records cannot be recovered from aggregate results, with a quantifiable privacy guarantee.
Digital worker: A packaged automation that performs a defined operational role across systems, often combining RPA, APIs, workflows, and AI. It is software, not an employee, and requires ownership and oversight. Browse 27,000+ ready-made workflows →
Direct prompt injection: Prompt injection delivered explicitly through a user's input, such as instructions to ignore policies, reveal protected context, or misuse tools. Browse the prompt library →
Distillation: Training a small model to imitate a larger one. Retains much of the quality at a fraction of the serving cost.
Document loader: A component that reads data from a source such as PDFs, websites, drives, databases, or APIs and converts it into a format suitable for processing and indexing.
Drift detection: Monitoring for shifts in input distribution or model output that signal a model has gone stale.
Durable execution: Execution that persists workflow progress so a process can recover after failures, crashes, or outages and continue from recorded state rather than restarting from the beginning. Browse 27,000+ ready-made workflows →

E

Error handling: Workflow branches that catch failures — retries, fallback paths, alert notifications — so one bad record does not silently kill a business process.
ETL: Extract, Transform, Load — moving data from sources, reshaping it, and writing it to a destination. Many automation workflows are lightweight ETL pipelines.
Embedding: A numeric vector representing the meaning of text, so similar ideas land near each other and can be searched by similarity rather than keywords.
Edge AI: Running models on local devices rather than in the cloud — lower latency and better privacy, tighter compute limits.
Edge inference: Running an AI model near the data source or on a local device rather than a centralized cloud. It can reduce latency and data transfer but is constrained by device resources and update management.
Emergence: Capabilities that appear only above a certain model scale. Disputed — some apparent jumps are artefacts of how the metric was measured.
Encoder: The half of a model that turns input into an internal representation. Used alone for embeddings and classification. Explore AI tools →
Encryption at rest: Protecting stored data using encryption so stolen disks, backups, or database files are not readable without keys. Key management is as important as the encryption algorithm.
Encryption in transit: Protecting data while it moves across networks, typically using TLS. It prevents passive interception but does not make an untrusted endpoint safe.
Ensemble: Combining several models and pooling their predictions. Reliably beats any single member, at multiplied inference cost.
Episodic memory: Memory of specific past interactions or events, including what happened and when. An agent can use episodic memory to avoid repeating failures or continue prior work. Explore the AI agents directory →
Epoch: One full pass over the training data. More epochs improve fit until they start memorising instead of generalising.
Error workflow: A dedicated workflow triggered when another workflow fails. It can capture diagnostics, notify owners, open an incident, quarantine data, or start an approved recovery process. Browse 27,000+ ready-made workflows →
EU AI Act: The European Union's risk-tiered AI regulation, imposing obligations that scale with an application's assessed risk.
Eval set: A fixed collection of test cases with expected outcomes, run against every change. The difference between knowing a prompt got better and hoping it did. Browse the prompt library →
Evaluation or eval: A systematic test of an AI system against defined tasks, data, metrics, and acceptance thresholds. Evaluate the entire application, not only the base model.
Event-driven automation: Automation that reacts when an event occurs, such as an order being paid. It is usually faster and more efficient than repeatedly checking whether something changed. Browse 27,000+ ready-made workflows →
Exact-match evaluation: A metric that counts an answer as correct only when it exactly matches the reference after defined normalization. It is strict and unsuitable when multiple phrasings are valid.
Exactly-once effect: A business outcome that occurs only once even when underlying messages or retries may occur multiple times. In distributed systems this is usually achieved through idempotency, deduplication, and transactional design.
Exception handling: The logic used to detect, classify, record, and respond to errors. Good exception handling distinguishes retryable failures, permanent validation errors, and cases requiring human intervention.
Exception rate: The percentage of cases that leave the normal automated path because of errors, missing data, low confidence, or unusual conditions. Exception cost can determine whether an automation is worthwhile. Browse 27,000+ ready-made workflows →
Excessive agency: A security risk in which an AI system has unnecessary functionality, permissions, or autonomy, allowing harmful actions when the model is manipulated or simply wrong.
Execution or run: One attempt to process a workflow from its trigger through completion, failure, cancellation, or pause. A run normally contains logs, inputs, outputs, timing, and status for each step. Browse 27,000+ ready-made workflows →
Executor: The component that performs planned tasks using tools or workflows and returns observable results. Separating planning from execution can improve control and debugging.
Expert system: A 1980s approach encoding human knowledge as explicit if-then rules. Brittle and expensive to maintain, which is why learned models displaced it.
Explainability: Being able to give a human-understandable account of why a model produced an output.
Exponential backoff: A retry strategy that increases the delay after each failure, often doubling it. This gives an overloaded service time to recover and reduces repeated pressure.
Extract, Transform, Load (ETL): A data pipeline pattern that extracts data from sources, transforms it into the required structure, and loads it into a destination. Many non-AI automations are lightweight ETL systems. Browse 27,000+ ready-made workflows →

F

Fair-code: A licensing model (used by n8n) where source code is open and free for most uses, but commercial hosting rights are restricted.
Fine-tuning: Further training a base model on your own examples to specialize its behavior. Heavier than prompting or RAG; used when consistency matters at scale.
F1 score: The harmonic mean of precision and recall. It is useful when both false positives and false negatives matter, but it hides the distinct tradeoff between them.
Faithfulness: Whether an output accurately reflects its provided source material without adding unsupported claims or contradicting evidence. The term is often used similarly to groundedness.
Fallback: An alternative path used when the preferred action, model, or service is unavailable or unreliable. Examples include a cheaper model, cached answer, rules-based path, or human review.
Fallback model: A secondary model used when the preferred model is unavailable, too slow, too costly, or fails quality checks. Fallback behavior must be tested because model capabilities differ.
Feature engineering: Deriving the input signals a model actually learns from. Still where most of the accuracy comes from in tabular problems.
Feature store: A shared service for computing and serving model inputs consistently between training and production.
Federated learning: Training across decentralised data without moving it, by sharing model updates instead of raw records.
Few-shot learning: Supplying a handful of worked examples in the prompt so the model infers the pattern. Usually the fastest quality improvement available. Browse the prompt library →
Few-shot prompting: Including a small number of input-output examples in the prompt to demonstrate the desired pattern, tone, or format. Examples should cover representative and difficult cases. Browse the prompt library →
Foundation model: A broadly trained model that can be adapted to many tasks through prompting, tools, retrieval, or fine-tuning rather than being built for only one narrow use case. Explore AI tools →
Fuzzy logic: Reasoning with degrees of truth rather than true/false. Used in control systems where thresholds should be gradual.

G

Guardrails: Constraints around AI systems — input validation, output filtering, human approval gates — that keep automated AI behavior inside acceptable bounds.
GPT (custom GPT): A configurable ChatGPT persona with custom instructions, knowledge files, and actions, shareable via OpenAI's GPT Store. ChatGPT GPT directory →
Garbage in, garbage out (GIGO): Output quality is capped by input quality. In AI it is worse than in classic software, because bad inputs produce confident, fluent, wrong answers.
Generative AI: AI that creates new content such as text, images, audio, video, or code by modeling patterns learned from training data.
Genetic algorithm: Optimisation that mimics natural selection — generate candidates, keep the best, mutate and recombine. Useful when there is no gradient to follow.
GET: An HTTP method for retrieving a resource without intending to change it. Properly designed GET requests are safe and should be repeatable.
Golden dataset: A curated set of representative examples with trusted expected answers, labels, or scoring criteria. It serves as a stable benchmark for comparing prompts, models, and system versions. Compare 298 models →
GPU: The processor type most AI training and inference runs on, because it performs many matrix operations in parallel.
Gradient descent: Repeatedly nudging parameters in the direction that most reduces error. 'Learning rate' is how big each nudge is.
GraphQL: A query language and API runtime in which clients request specific fields through a typed schema. It reduces over-fetching but introduces query complexity and cost-control considerations.
Greedy decoding: Always taking the single most likely next token. Deterministic and often flat or repetitive.
Green AI: Reducing the energy and carbon cost of training and serving models, including by choosing smaller models.
Groundedness: The degree to which an answer's claims are supported by the supplied context or authoritative evidence. Groundedness should be assessed at the claim level when possible.
Grounding: Connecting a model's response to authoritative external information such as documents, databases, search results, or tool outputs. Grounding improves verifiability and freshness.
Guardrail: A technical or procedural control that constrains AI inputs, outputs, tool use, permissions, or actions. Guardrails reduce risk but are layers of defense, not guarantees.

H

Hallucination: When an AI model states something false with confidence. Mitigated with retrieval (RAG), citations, and verification steps in your workflow.
Human-in-the-loop: Workflow designs where a person reviews or approves AI output before it takes effect. The standard pattern for high-stakes automations.
Halt condition: The explicit rule for when an agent stops — goal met, budget spent, or confidence too low. Explore the AI agents directory →
Heuristic search: Searching a space of possibilities using a rule of thumb to decide what to explore next, trading guaranteed optimality for tractable runtime.
Hierarchical agent: A supervising agent delegating subtasks to specialised sub-agents. Explore the AI agents directory →
HMAC signature: A cryptographic code generated from a message and shared secret. Receivers recompute it to verify that a webhook or request is authentic and unchanged. Browse 27,000+ ready-made workflows →
HTTP header: Metadata attached to a request or response, used for content type, authorization, caching, signatures, rate limits, tracing, and other protocol behavior.
HTTP method: A verb expressing the intended operation on a resource, commonly GET, POST, PUT, PATCH, or DELETE. The same endpoint may support different methods with different behavior.
HTTP request: A message sent by a client to a server containing a method, URL, headers, and sometimes a body. Workflow tools construct requests to call external APIs. Browse 27,000+ ready-made workflows →
HTTP response: A server's reply containing a status code, headers, and usually a body. Automations should validate both the status and the response content rather than assuming success. Browse 27,000+ ready-made workflows →
HTTP status code: A three-digit result code returned by an HTTP server. 2xx indicates success, 4xx usually indicates a request or authorization problem, and 5xx indicates a server-side failure.
Human escalation rate: The percentage of AI or workflow cases referred to a person. A healthy rate balances automation value against the risk of forcing uncertain cases through automatically. Browse 27,000+ ready-made workflows →
Human evaluation: Assessment by people using a defined rubric. Human evaluation captures nuance but is slower, more expensive, and subject to disagreement, fatigue, and evaluator bias.
Human-on-the-loop (HOTL): A design in which automation operates independently while a person monitors it and can intervene. It offers less friction than pre-approval but requires strong observability and rapid stop controls. Browse 27,000+ ready-made workflows →
Human-out-of-the-loop: A design in which no person reviews individual decisions or actions. It is appropriate only for sufficiently low-risk, bounded, reversible, and well-tested operations.
Hybrid search: Combining semantic and keyword retrieval, often followed by score fusion or reranking. Hybrid search commonly performs better than either method alone across mixed query types.
Hyperautomation: A strategy of systematically discovering, prioritizing, automating, measuring, and improving as many suitable business processes as possible using tools such as process mining, APIs, RPA, AI, and orchestration. Browse 27,000+ ready-made workflows →
Hyperparameter: A setting chosen before training rather than learned during it: learning rate, batch size, layer count. Tuning these is most of practical ML work.

I

iPaaS: Integration Platform as a Service — cloud platforms (Zapier, Make, Workato) that host your integrations and automations without you running servers.
Idempotency: Designing steps so running them twice has the same effect as once. Critical for automations that may retry after partial failures (e.g., avoid double-charging).
Inference: Running a trained model to produce output. Inference cost and latency — not training — dominate most production AI budgets.
Improper output handling: Using model-generated output in code, queries, HTML, shell commands, or downstream actions without appropriate validation, sanitization, or authorization.
Indexing pipeline: The process that collects, cleans, segments, enriches, embeds, and stores documents for retrieval. It should also handle updates, deletion, permissions, and failed documents.
Indirect prompt injection: Malicious instructions hidden in external content the model reads, such as a webpage, email, document, or tool result. Retrieved content must not automatically gain instruction authority. Browse the prompt library →
Inference cost: What it costs to serve predictions. Usually dwarfs training cost over a product's life, and is the number that decides unit economics.
Input token: A token sent to the model as instructions, conversation history, retrieved context, or tool results. Input tokens consume context capacity and are often billed separately.
Instruction hierarchy: The precedence rules used when instructions conflict, typically placing platform or system instructions above developer and user instructions. Retrieved content should usually be treated as data, not authority.
Instruction tuning: Fine-tuning on examples formatted as instruction-and-response so the model reliably does what it is asked rather than merely continuing text. Explore AI tools →
Integration platform as a service (iPaaS): A cloud platform for connecting applications and automating data flows without operating the underlying integration infrastructure. Examples commonly include visual builders, managed connectors, monitoring, and governance. Browse 27,000+ ready-made workflows →
Intelligent automation: The combination of conventional automation with AI capabilities such as classification, extraction, prediction, or language understanding. The deterministic workflow controls the process while AI handles unstructured decisions. Browse 27,000+ ready-made workflows →
Interpretability: Understanding a model's internal mechanics, as distinct from explaining a single decision after the fact.

J

Jailbreak: A form of prompt injection intended to bypass a model's safety or policy constraints. Filters and system prompts alone cannot guarantee resistance. Browse the prompt library →
Jitter: A small random variation added to retry delays so many workers do not retry at exactly the same moment. Jitter helps prevent a synchronized thundering-herd problem.
JSON: JavaScript Object Notation, a widely used text format for structured data made of objects, arrays, strings, numbers, booleans, and null values.
JSON mode: A provider setting guaranteeing syntactically valid JSON. Guarantees parseability, not that the fields are correct.
JSON Schema: A vocabulary for describing and validating the structure, types, required fields, and constraints of JSON data. It is commonly used to define structured model outputs and tool parameters.

K

K-means: A clustering method that partitions data into k groups around centroids. Requires you to pick k, and assumes roughly spherical clusters.
Keyword search: Search based on matching words or lexical signals. It is strong for exact names, codes, and rare terms that embedding search may overlook. Explore AI tools →
Knowledge base: An organized collection of information used for search, retrieval, support, or reasoning. A useful knowledge base includes ownership, permissions, versioning, and freshness processes.
Knowledge graph: Data modelled as entities and the relationships between them. Good at answering questions that require joining facts, which is exactly where plain text retrieval struggles.
Knowledge representation: How facts and relationships are encoded so software can reason over them — ontologies, graphs, logical statements.

L

LLM (Large Language Model): A neural network trained on large text corpora to generate and reason over language — GPT, Claude, Gemini, Llama. Compare current models on live benchmark data. AI model comparison →
LoRA: Low-Rank Adaptation — a small add-on file that customizes an image or language model (a face, a style, a domain) without retraining the whole model.
Low-code: Visual building plus small amounts of code where it counts — an n8n Code node, a Make custom function. The pragmatic middle ground.
Large language model (LLM): A model trained on large amounts of text and code to predict and generate token sequences. LLMs can perform many language tasks but do not guarantee truth or deterministic reasoning. Explore AI tools →
Latency: The elapsed time between a request or trigger and its response or outcome. AI latency includes queueing, retrieval, model inference, tool calls, and output processing. Browse 27,000+ ready-made workflows →
Latency budget: The maximum response time allocated to an end-to-end request and divided among retrieval, model calls, tools, validation, and network overhead.
Least privilege: Granting each workflow, user, or agent only the permissions required for its current task and no more. Least privilege reduces the impact of mistakes and attacks. Explore the AI agents directory →
LLM-as-a-judge: Using a language model to score another model's output against a rubric. It scales qualitative evaluation but can introduce judge bias, inconsistency, and preference for certain styles. Explore AI tools →
LLM-as-judge: Using one model to grade another's output against a rubric. Scales evaluation far beyond human review, and inherits the judge's biases. Explore AI tools →
Load balancing: Distributing requests across multiple servers, workers, models, or regions to improve capacity and availability. The strategy may consider health, latency, cost, or session state.
Local MCP server: An MCP server running on the same device or environment as the host, commonly connected through standard input and output. It can access local resources within its granted permissions. Explore MCP servers →
Log: A timestamped record of an event or diagnostic detail. Logs should be structured, correlated, access-controlled, retained intentionally, and scrubbed of secrets and unnecessary personal data.
Long-running workflow: A process that may remain active for hours, days, or months while waiting for events, approvals, deadlines, or external systems. Durable state and safe resumption are essential. Browse 27,000+ ready-made workflows →
Long-term memory: Information persisted across sessions, such as preferences, facts, or prior outcomes. Long-term memory requires rules for consent, relevance, correction, expiration, and access control.
Loop or iteration: A control structure that repeats steps for each item or until a condition is met. Loops require limits and error handling to prevent runaway executions.
Loss function: The number training tries to minimise — the formal statement of what 'wrong' means. Choosing it badly optimises the wrong thing perfectly.
Low-code automation: Visual workflow development supplemented by small amounts of code for transformations, custom APIs, or specialized logic. It provides more flexibility than no-code while retaining faster development than fully custom software. Browse 27,000+ ready-made workflows →
Low-Rank Adaptation (LoRA): A parameter-efficient fine-tuning method that learns low-rank updates to selected model weights. LoRA produces relatively small adapter files rather than a complete new model. Explore AI tools →

M

Multi-agent system: Several specialized AI agents cooperating on one objective — e.g., a researcher, a writer, and a reviewer passing work between them.
MCP (Model Context Protocol): An open protocol for connecting AI assistants to external tools and data sources through standardized servers. Full explanation →
Multimodal model: A model that handles more than text — images, audio, or video in and/or out (GPT-4o, Gemini, Claude with vision).
Marketplace commission: The platform fee on each sale. On Neura Market sellers keep 80–90% of marketplace sales and 95% via direct links. Seller pricing →
Machine learning: Building software that learns patterns from examples instead of being explicitly programmed. You supply data and a target; the training process finds the rules.
Maintenance burden: The ongoing work required to keep an automation reliable as APIs, schemas, models, credentials, rules, and business processes change. Browse 27,000+ ready-made workflows →
Manual trigger: A user-initiated start mechanism used for testing, one-off jobs, or processes requiring deliberate initiation. Manual triggers are usually not appropriate for unattended production workflows. Browse 27,000+ ready-made workflows →
Maximum output tokens: A hard limit on how many tokens the model may generate. It controls worst-case cost and latency but can truncate an answer if set too low.
MCP client: The protocol component created by a host to maintain a dedicated connection with an MCP server and exchange requests, capabilities, and results. Explore MCP servers →
MCP host: The AI application that manages the user experience and coordinates one or more MCP clients, such as an assistant or development environment. Explore MCP servers →
MCP prompt: A reusable prompt template exposed by an MCP server to help users or models perform a defined task consistently. Explore MCP servers →
MCP resource: File-like or structured data exposed by an MCP server for a client to read, such as document contents, database records, or API responses. Explore MCP servers →
MCP server: A local or remote program that exposes capabilities to AI applications through MCP. A server may provide tools, resources, prompts, or interactive applications. Explore MCP servers →
MCP tool: A callable function exposed by an MCP server that lets an AI system perform an operation. Tool parameters and outputs should have clear schemas and appropriate approval controls. Explore MCP servers →
MCP transport: How an MCP client and server exchange messages, typically stdio for local servers or HTTP for remote ones. Explore MCP servers →
Metadata filtering: Restricting retrieval using structured attributes such as customer, date, document type, permissions, or region. Filtering improves precision and helps enforce access control.
Metric: A numeric measurement tracked over time, such as latency, error rate, task success, token usage, or human escalation rate. Useful metrics connect technical behavior to user outcomes.
Model: The trained artefact — a set of learned numbers plus the code to run them. A model is a frozen snapshot of what training found; it does not learn from you as you use it.
Model card: A published summary of a model's intended use, training data, evaluations and known limits.
Model chaining: Passing output from one model into another, often a cheap model for triage and an expensive one for the hard cases.
Model collapse: Degradation that occurs when models are trained repeatedly on AI-generated output, narrowing diversity each generation.
Model Context Protocol (MCP): An open protocol that standardizes how AI applications connect to external tools, resources, and prompt templates through a client-server architecture. Explore MCP servers →
Model drift: A decline or change in system performance as data, users, providers, or real-world conditions evolve. Continuous measurement helps distinguish drift from isolated failures.
Model gateway: An interface that presents multiple AI models through a common API and routes requests among providers. It simplifies switching but may hide provider-specific features or behavior.
Model registry: A catalogue of model versions with their metadata, so you can tell which version produced which result.
Model routing: Selecting a model dynamically based on task type, risk, quality requirements, latency, cost, context size, or modality. Routing can improve economics but adds evaluation and fallback complexity.
Multi-hop reasoning: Answering a question that requires combining several separate facts. Where single-shot retrieval usually fails.
Multimodal AI: Models handling more than one input or output type — text plus images, audio or video.

N

Node: In flow-based tools like n8n, one block in the visual graph: a trigger, an app action, or a logic step. Workflows are wired-together nodes.
Negative prompt: In image generation, the list of things you do NOT want (blur, extra fingers, watermarks). As influential as the positive prompt.
No-code: Building software behavior through visual interfaces instead of programming. Most automation platforms are no-code with escape hatches to code.
n8n workflow: An automation built in n8n as a graph of connected nodes, stored as a single JSON document. One trigger node starts it and each subsequent node performs one operation on the items passed to it. Full explanation →
Natural language generation (NLG): Producing human-readable text from data or instructions.
Natural language processing (NLP): The field of getting software to work with human language. Largely absorbed into language models. Explore AI tools →
Neura Market AI Automation Glossary: A free plain-English reference for AI agents, LLMs, MCP, RAG, workflow automation, integrations, and production concepts. Explore the continually updated glossary at https://www.neura.market/glossary Explore MCP servers →
Neural network: A model built from layers of simple numeric units whose connection strengths are adjusted during training. The 'depth' is just how many layers sit between input and output.
No-code automation: Building workflows through visual configuration without writing conventional code. It speeds up common integrations but can become limiting when logic, scale, testing, or security requirements grow. Browse 27,000+ ready-made workflows →
Normalization: Converting inconsistent data into a standard representation, such as canonical phone numbers, lowercase email addresses, or standardized country codes.

O

Open-weight model: A model whose trained weights are downloadable (Llama, Mistral, DeepSeek), so you can run it on your own hardware.
OAuth 2.0: A framework that allows an application to obtain limited access to a user's resources without receiving the user's password. It uses grants, scopes, access tokens, and often refresh tokens.
Observability: The ability to understand a system's internal behavior from outputs such as traces, logs, metrics, and evaluation results. AI observability should include prompts, retrieval, model versions, tools, costs, and outcomes with sensitive data protected.
Offline evaluation: Evaluation performed on a fixed dataset before deployment. It is safe and repeatable but may not represent real traffic, user behavior, or production tool failures.
Online evaluation: Evaluation using live or shadow production traffic. It reveals real-world performance but requires careful experiment design, monitoring, privacy controls, and rollback capability.
Ontology: A formal description of the concepts in a domain and how they relate. The vocabulary a knowledge graph is built on.
Orchestration: Central coordination of multiple services, agents, or workflow steps by a controller that decides what runs, in what order, and how failures are handled. Browse 27,000+ ready-made workflows →
Output token: A token generated by the model. Output tokens are frequently more expensive and slower than input tokens, making concise output controls valuable.
Overfitting: A model that has memorised its training data and performs worse on anything new. Detected by a train/validation gap, not by training accuracy.

P

Polling: Checking a service for new data on a schedule (e.g., every 5 minutes). Simpler than webhooks but slower and rate-limit-hungry.
Prompt: The instruction text you give an AI model. Prompt quality drives output quality, which is why curated prompt libraries exist. Browse prompts →
Prompt engineering: Systematically designing prompts — role framing, examples, output formats, constraints — to get reliable results from LLMs. Full explanation →
Prompt pack: A curated bundle of prompts for one job (e.g., SEO audits, cold email) designed to work together. Browse prompt packs →
Pagination: Retrieving a large result set in smaller pages using offsets, cursors, or continuation tokens. Production workflows must continue until the final page and guard against duplicate or missing records.
Parallel execution: Running independent steps or items at the same time to reduce total duration. Parallelism improves throughput but increases concurrency, rate-limit, ordering, and race-condition risks.
Parameter: One learned number inside a model. Parameter count ('7B', '70B') is a rough proxy for capacity, not for quality — a well-trained small model routinely beats a badly-trained large one.
PATCH: An HTTP method used to apply a partial update to a resource. A PATCH request should specify only the fields or operations being changed.
Pattern recognition: Identifying regularities in data — the general problem underneath classification, detection and clustering.
Payload: The substantive data carried in a request, event, or message. A payload should be validated against an expected schema before downstream use.
Perplexity: How surprised a language model is by text — lower means the text is more predictable to it. A training diagnostic, not a quality score. Explore AI tools →
Personally identifiable information (PII): Information that identifies or can reasonably be linked to a person. Examples include names combined with contact, financial, location, health, or government identification data.
Plan-and-execute: Producing a full plan first, then executing steps. More predictable and auditable than deciding each step on the fly.
Plan-and-execute pattern: An agent pattern that first creates a multi-step plan and then executes or revises it. It improves visibility on complex tasks but can waste work if the initial plan is poor. Explore the AI agents directory →
Planner: The component or agent that decomposes a goal into tasks, dependencies, and success criteria. Planning should remain grounded in available tools, permissions, time, and budget. Explore the AI agents directory →
POST: An HTTP method commonly used to create a resource or start an operation. Repeating a POST can create duplicates unless the API supports idempotency.
Pre-training: The expensive first stage where a model learns general structure from a large corpus, before any task-specific tuning.
Precision: Of the items a system labeled positive or retrieved, the proportion that were actually relevant or correct. High precision means fewer false positives.
Prefill: The phase where the model reads your prompt before generating. Usually billed differently from output and is what prompt caching accelerates. Browse the prompt library →
Prefix cache: Reusing the computed state of a repeated prompt prefix across calls. Large cost and latency saving for long, stable system prompts. Browse the prompt library →
Pretraining: Large-scale initial training on broad datasets that teaches a foundation model general patterns and capabilities before task-specific alignment or adaptation. Explore AI tools →
Priority queue: A queue that processes higher-priority jobs before lower-priority ones. Priority must be designed carefully so routine work is not starved indefinitely.
Probabilistic automation: Automation that uses models whose outputs may vary or contain uncertainty. It is appropriate for language, images, classification, and judgment tasks, but should include confidence thresholds, validation, and fallbacks. Browse 27,000+ ready-made workflows →
Procedural memory: Retained know-how — the steps for doing something, as opposed to facts about it.
Process discovery: The structured identification and documentation of how work is performed, including inputs, decisions, exceptions, systems, owners, and outcomes, before automation is designed. Browse 27,000+ ready-made workflows →
Process mining: Analyzing event logs from business systems to reconstruct how processes actually operate, identify bottlenecks and variation, and find automation opportunities. Browse 27,000+ ready-made workflows →
Prompt caching: Reusing processing for repeated prompt prefixes or contexts so later model calls are faster or cheaper. Cache behavior and pricing vary by provider. Browse the prompt library →
Prompt injection: An attack or failure mode in which untrusted input changes a model's behavior or causes it to follow instructions that conflict with the application's intent. Browse the prompt library →
Prompt template: A reusable prompt structure with placeholders filled at runtime. Templates improve consistency and version control but still require validation of inserted data. Browse the prompt library →
Prompt versioning: Storing prompts as versioned artifacts with change history, owners, test results, and rollback capability. Production prompts should be managed like code or configuration. Browse the prompt library →
Provenance: A verifiable record of where a piece of content came from and how it was altered.
PUT: An HTTP method commonly used to replace a resource at a known location. PUT is generally expected to be idempotent.

Q

Quantization: Storing weights at lower numeric precision to cut memory and speed up inference, accepting a small accuracy loss. What makes local models fit on consumer hardware.
Quantum computing: Computing with quantum states rather than binary bits. Largely unrelated to current AI in practice, despite frequent co-marketing.
Queue: A buffer that stores work until a consumer can process it. Queues absorb traffic spikes, decouple systems, and enable retries, prioritization, and controlled concurrency.
Queue mode: An execution architecture in which a main service accepts and schedules work while separate workers process queued jobs. It supports horizontal scaling and isolates workload spikes.
Quota: A provider-defined allowance for requests, tokens, storage, spending, or concurrency over a period. Unlike momentary rate limits, quotas often require waiting, upgrading, or requesting an increase.

R

Rate limit: The maximum number of API calls a service allows per time window. Well-built workflows batch requests and back off when limits approach.
RAG (Retrieval-Augmented Generation): Fetching relevant documents (usually via a vector database) and injecting them into the prompt, so the model answers from your data instead of memory. Full explanation →
Reasoning model: An LLM variant that spends extra compute "thinking" before answering, trading latency and cost for better multi-step problem solving.
Race condition: A defect in which the outcome depends on the unpredictable timing of concurrent operations. Locks, atomic updates, queues, and idempotent designs can reduce race conditions.
Random seed: A value intended to make sampling more repeatable when supported. Model updates, infrastructure, and nondeterministic operations can still prevent perfectly identical results.
Rate limiting: Provider-imposed caps on request or token volume. Designing for them is what separates a demo from a production integration. Browse 27,000+ ready-made workflows →
ReAct: An agent pattern interleaving reasoning with tool calls, so each action is justified by an explicit thought. Explore the AI agents directory →
ReAct pattern: An agent pattern that interleaves reasoning and actions so the model chooses a tool, observes the result, and adapts its next step. The name combines reasoning and acting. Explore the AI agents directory →
Recall: Of all relevant or positive items that existed, the proportion the system successfully found. High recall means fewer false negatives.
Recommendation engine: A system suggesting items based on behaviour and similarity. The most commercially proven form of applied ML.
Recursive prompting: Feeding a model's output back as input to refine it across passes.
Red team: Deliberately attacking your own system to find failures before users or attackers do.
Red teaming: Structured adversarial testing intended to uncover harmful behavior, security weaknesses, abuse paths, and unexpected failures before attackers or users find them.
Redaction: Removing or masking sensitive content before it enters logs, prompts, model providers, or downstream systems. Redaction should be tested because formats and identifiers vary widely.
Reflection: A process in which an agent reviews its result or trajectory and decides whether to revise it. Reflection can improve quality but adds cost and may reinforce errors without external evidence. Explore the AI agents directory →
Reflexion: Having an agent critique its own previous attempt and retry with that critique in context. Explore the AI agents directory →
Refresh token: A longer-lived credential used to obtain new access tokens after they expire. Refresh tokens require stronger storage protection and revocation controls.
Regression: Predicting a continuous number rather than a category.
Regression test: A repeated test designed to detect whether a change broke behavior that previously worked. AI regression tests should allow for valid variation while protecting critical requirements.
Regularization: Techniques that penalise complexity to curb overfitting — weight decay, dropout, early stopping.
Reinforcement learning: Learning by acting and receiving rewards rather than from labelled answers. Powerful where you can simulate cheaply, awkward where you cannot.
Reinforcement learning from human feedback (RLHF): A model-alignment approach that uses human preference signals to reward more desirable outputs. Related methods may use AI feedback or direct preference optimization.
Remote MCP server: An internet-hosted MCP server accessed by compatible clients. Remote servers require authentication, secure transport, explicit trust, and careful review of exposed capabilities. Explore MCP servers →
Remote procedure call (RPC): A method of invoking a function on another system as if it were local. RPC systems emphasize named operations rather than REST-style resources.
Reranking: Applying a stronger relevance model to reorder an initial set of retrieved candidates. Reranking improves quality while avoiding the cost of scoring the entire collection with the stronger model.
Responsible AI: Practices covering fairness, transparency, privacy and accountability across an AI system's life.
REST: An API style centered on resources addressed by URLs and manipulated through standard HTTP methods. REST is a design approach, not a single protocol or data format.
Resume: Continuing a paused or failed process from saved state rather than starting again. A safe resume mechanism must know which actions already produced external effects.
Retrieval: Selecting potentially relevant information from a collection for a query. Retrieval quality depends on document processing, query formulation, indexing, filters, ranking, and freshness.
Retrieval top-k: The number of highest-ranked candidates returned by a retriever. A larger k improves recall but can add irrelevant context, latency, and token cost.
Retrieval-Augmented Generation (RAG): A pattern that retrieves relevant information at request time and includes it in the model context before generation. RAG improves access to private or current knowledge without retraining the model. Explore AI tools →
Retry: A new attempt after a temporary failure. Retries should target transient errors, use limits and delays, and avoid repeating non-idempotent actions without protection.
Robotic process automation (RPA): Software bots that imitate human interactions with user interfaces, such as clicking buttons or copying fields. RPA is useful when an application lacks an API but is more brittle than API-based automation. Browse 27,000+ ready-made workflows →
Role-based access control (RBAC): Assigning permissions to roles and then assigning users or services to those roles. RBAC simplifies administration but should be combined with least privilege and periodic review.
ROUGE: Overlap-based scoring for summarisation. Same trade-off as BLEU — measures wording, not meaning.
Rubric: The explicit criteria a grader scores against. Writing one is usually what reveals that 'good output' was never defined.

S

Scenario: Make's (formerly Integromat) name for a workflow: a visual sequence of modules that runs on a trigger or schedule.
Self-hosted automation: Running the automation platform on your own infrastructure (typical with n8n or Activepieces) for data control and lower per-task costs, at the price of maintenance. n8n workflow templates →
System prompt: The hidden instruction that sets an AI assistant's behavior, tone, and rules before any user input arrives.
Stable Diffusion checkpoint: A trained image-generation model file. Different checkpoints specialize in styles: photorealism, anime, product shots. Stable Diffusion directory →
Saga pattern: A method for coordinating a multi-service transaction as a series of local transactions with compensating actions. It provides eventual consistency when a single atomic transaction is impossible.
Sandbox: An isolated environment that limits what code, tools, or agents can access or affect. Sandboxes reduce blast radius but still require resource limits, monitoring, and secure configuration.
Sandboxing: Running agent-executed code in an isolated environment with restricted filesystem, network and credential access. Explore the AI agents directory →
Schedule-based automation: Automation launched at defined times or intervals, such as every hour or each Monday at 8 a.m. It is appropriate for reports, synchronization, cleanup, and systems that cannot emit events. Browse 27,000+ ready-made workflows →
Schema: A formal description of expected data fields, types, constraints, and nesting. Schemas make integrations safer by enabling validation and predictable mapping. Browse 27,000+ ready-made workflows →
Scratchpad: Working notes a model writes for itself mid-task. Improves multi-step accuracy and gives you something to inspect when it goes wrong.
Secret: Sensitive configuration such as an API key, token, password, or signing key. Secrets should be stored outside workflow code, access-controlled, rotated, and excluded from logs. Browse 27,000+ ready-made workflows →
Self-consistency: Sampling several answers and taking the most common. A cheap accuracy win on problems with one right answer.
Self-critique: A model-generated assessment of its own output against criteria. It is useful as one signal but should not replace independent tests or human review for high-stakes work.
Semantic cache: A cache that reuses a prior result when a new request is meaningfully similar rather than textually identical. It requires thresholds, freshness limits, and protection against cross-user data leakage.
Semantic chunking: Splitting documents on meaning rather than fixed character counts, so retrieved passages stay coherent.
Semantic memory: Stored generalized knowledge or facts separated from the exact event in which they were learned. It is often represented through structured records or searchable embeddings. Explore AI tools →
Semantic search: Search that uses embeddings to find items with similar meaning even when they do not share exact keywords. It is useful for paraphrases and conceptual matches. Explore AI tools →
Semantic web: The effort to publish web data with machine-readable meaning, via standards like RDF and schema.org. Structured data markup is its most widely-used surviving piece.
Semantic-similarity evaluation: A metric that estimates whether an output has similar meaning to a reference, often using embeddings or a model. Similar meaning does not guarantee factual correctness. Explore AI tools →
Sensitive information disclosure: Exposure of private, confidential, regulated, or proprietary data through model responses or application behavior. Prevention includes minimization, access control, redaction, and output review.
Sentiment analysis: Classifying the attitude expressed in text. Reliable in aggregate, unreliable on sarcasm and short strings.
Sequential execution: Running steps one after another so each step can depend on the previous result. It is simpler and safer when order matters but is slower than parallel execution.
Serialization: Converting an in-memory object into a storable or transmittable format such as JSON. Serialization lets data cross process or network boundaries.
Service-level agreement (SLA): A formal commitment between a provider and customer defining service expectations and possible remedies. An SLA is contractual; an SLO is an engineering target.
Service-level objective (SLO): An internal reliability target for a service, such as 99.9% successful executions or 95% of runs finishing within 30 seconds.
Shadow AI: Employees using AI tools outside sanctioned channels, moving company data through unreviewed services.
Shadow deployment: Running a new model alongside the live one on real traffic without acting on its output, to compare safely.
Short-term memory: Information retained within a current conversation, thread, or workflow run. It supports continuity but must be summarized or trimmed as context grows. Browse 27,000+ ready-made workflows →
Similarity score: A numeric measure of closeness between a query and candidate. Scores are model- and index-specific and should be calibrated rather than treated as universal probabilities.
Software development kit (SDK): A language-specific library that wraps an API with convenient classes and functions. SDKs speed development but can lag behind the underlying API or hide important behavior.
Span: One timed unit inside a trace, such as a single tool call.
Specialist agent: An agent configured for a narrow role, domain, or tool set. Specialization can improve performance and reduce permissions compared with one general-purpose agent. Explore the AI agents directory →
Speech recognition: Converting spoken audio to text. Also called ASR or speech-to-text.
Standard input/output transport (stdio): A local MCP transport in which a client communicates with a server process through its standard input and output streams. Explore MCP servers →
State: The information a workflow retains about its current position, prior results, pending work, and decisions. State lets a process pause and later continue without starting over. Browse 27,000+ ready-made workflows →
Stateful workflow: A workflow that persists information across steps, pauses, retries, or sessions. Stateful designs support long-running processes and agent memory but require careful storage and lifecycle management. Explore the AI agents directory →
Stateless workflow: A workflow that does not retain information between requests or runs. Each execution contains everything it needs, making scaling easier but limiting long-lived interactions. Browse 27,000+ ready-made workflows →
Step: One unit of work in a workflow. A step may call an application, run code, evaluate a condition, pause for approval, or invoke an AI model. Browse 27,000+ ready-made workflows →
Step budget: A hard cap on agent iterations. The main defence against a loop that never terminates. Explore the AI agents directory →
Stochastic parrot: A critique framing language models as recombining training text without understanding. Shorthand for the argument that fluency is not comprehension. Explore AI tools →
Stop sequence: A specified text pattern that tells a model to stop generating when encountered. Stop sequences help delimit outputs but can trigger unexpectedly if the sequence appears naturally. Browse 27,000+ ready-made workflows →
Straight-through processing (STP): Completing a process from start to finish without manual handling. STP rate is the percentage of cases completed automatically without exception or intervention.
Streamable HTTP transport: An MCP transport for communicating with remote servers over HTTP while supporting streamed messages. It replaces older remote transport patterns in current MCP architecture. Explore MCP servers →
Streaming: Returning tokens as they are produced rather than waiting for the whole response. Improves perceived speed; complicates error handling mid-response.
Streaming response: Delivering model output incrementally as it is generated. Streaming improves perceived latency but complicates validation because the full output is not yet available.
Strong AI: The hypothetical general, self-directed intelligence contrasted with narrow AI. A philosophical category, not an engineering roadmap.
Structured output: Model output constrained to a machine-readable structure such as validated JSON. Structured output improves parsing reliability but does not guarantee that the values are factually correct.
Subgoal: An intermediate objective an agent sets on the way to the main goal. Explore the AI agents directory →
Subworkflow: A reusable workflow called by another workflow to perform a focused function. Subworkflows reduce duplication, improve testing, and let teams change shared logic in one place. Browse 27,000+ ready-made workflows →
Supervised fine-tuning (SFT): Fine-tuning on labeled input-output examples that demonstrate desired responses. Data quality and representativeness are usually more important than raw example count. Explore AI tools →
Supervised learning: Training on labelled examples where the correct answer is supplied. Most business ML is supervised, and its cost is dominated by labelling.
Supervisor agent: An agent that assigns tasks, monitors specialists, resolves conflicts, and synthesizes results. The supervisor should have explicit limits and should not become an unobservable single point of failure. Explore the AI agents directory →
Swarm: Many lightweight agents working in parallel with little central control. Explore the AI agents directory →
Synthetic data: Machine-generated training data. Useful where real data is scarce or sensitive; risks amplifying the generating model's own blind spots.
System prompt leakage: Exposure of hidden system instructions or configuration. System prompts should not contain secrets, and applications should remain secure even if prompt contents become known. Browse the prompt library →

T

Trigger: The event that starts an automation — a new email, a form submission, a webhook call, or a schedule. Every workflow begins with exactly one trigger.
Task / Operation: The billing unit of most automation platforms: one executed action step. Pricing comparisons between Zapier, Make, and n8n usually come down to cost per 1,000 tasks. n8n vs Zapier comparison →
Token: The unit LLMs read and bill by — roughly ¾ of an English word. API pricing is quoted per million input and output tokens.
Temperature: A sampling setting controlling output randomness: low values give consistent, deterministic answers; high values give more varied, creative ones.
Tool use / function calling: Letting an LLM invoke defined functions (search, database queries, API calls) and use the results, turning a chat model into a doer.
Task mining: Capturing and analyzing user interactions on desktops to understand repetitive tasks. It can reveal automation opportunities but requires strong privacy and workforce-governance controls. Browse 27,000+ ready-made workflows →
Task or operation: A platform-specific billing or execution unit, commonly representing one action performed for one item. Because platforms count tasks differently, task-based prices should be compared using the same real workflow. Browse 27,000+ ready-made workflows →
Task success rate: The proportion of attempts that achieve the defined end-to-end goal. Success criteria should reflect the real business outcome, not merely a fluent model response.
Technological singularity: A speculated point at which self-improving AI advances beyond human ability to follow. Not a measurable engineering milestone.
Test case: A defined input, expected behavior, and assessment method used to verify a system. Strong test suites include normal, edge, adversarial, multilingual, and failure scenarios.
Throttling: Deliberately controlling request speed or concurrency to remain within capacity or rate limits. Throttling protects both the caller and the downstream service.
Throughput: The amount of work completed per unit of time, such as executions per minute. Throughput depends on concurrency, processing time, rate limits, and downstream capacity.
Time to value: The elapsed time from selecting an automation opportunity to realizing measurable benefits. Templates and managed connectors can reduce build time, while complex governance or poor process design can extend it. Browse 27,000+ ready-made workflows →
Timeout: The maximum time a step may wait before being treated as failed. Timeouts prevent hung requests from consuming resources indefinitely and should reflect the operation's realistic latency.
Token budget: A limit on input, reasoning, and output tokens allowed for a task. Token budgets control cost and context growth while forcing intentional information selection.
Tokenizer: The algorithm that converts text into token identifiers and back. Different models tokenize the same text differently, affecting cost and context usage.
Tool call: A model's request to run a named function with structured arguments. The mechanism by which a model does anything beyond producing text.
Tool result: Data returned to the model after a tool executes. Tool results should be validated, size-limited, and treated as potentially untrusted context.
Tool routing: Directing a request to a specific tool or tool group using rules, classifiers, embeddings, or an LLM. Routing reduces unnecessary tool exposure and improves specialization. Explore AI tools →
Tool schema: The typed description of a tool's name, purpose and arguments. The model only knows what this says, so a vague description produces wrong calls.
Tool selection: Choosing the most appropriate available capability for a task. Good tool selection considers permissions, reliability, latency, cost, and whether a tool is needed at all.
Tool use: Allowing a model or agent to invoke external capabilities such as search, databases, calculators, code execution, or APIs. Tools extend capability but also expand the security and failure surface. Explore the AI agents directory →
Tool-call accuracy: How often an agent selects the correct tool and supplies valid arguments at the appropriate time. It should include unnecessary and dangerous calls, not only successful syntax. Explore the AI agents directory →
Top-k: Restricting sampling to the k most likely next tokens.
Top-p (nucleus sampling): Sampling only from the smallest set of tokens whose probabilities sum to p. Controls variety more smoothly than a hard cutoff.
Top-p sampling: A sampling method that considers the smallest set of tokens whose cumulative probability reaches a threshold. It controls diversity differently from temperature and should not be treated as a confidence setting.
Total cost of ownership (TCO): The complete lifecycle cost of an automation, including licenses, usage, development, infrastructure, monitoring, maintenance, security, training, and human review. Browse 27,000+ ready-made workflows →
Trace: An end-to-end record of a request across workflow steps, model calls, retrieval, and tools. Traces reveal sequence, timing, inputs, outputs, and failure points. Browse 27,000+ ready-made workflows →
Training: The process of adjusting model parameters using data and an optimization objective. Training creates or modifies model behavior; inference uses the resulting model.
Training data: The examples a model learned from. Its coverage and biases become the model's coverage and biases, which is why data provenance matters more than architecture for most failures.
Transfer learning: Reusing a model trained on one task as the starting point for another. Why fine-tuning a foundation model beats training from scratch for almost everyone. Explore AI tools →
Transformer: The neural architecture behind essentially every modern language model, built on attention. Its key property is processing a whole sequence in parallel rather than left to right. Explore AI tools →
Tree-of-thought: Exploring several reasoning branches and selecting among them, instead of committing to one chain.
Trigger-to-action latency: The time between an originating event and the resulting automated action. It includes event-delivery delay, queueing, processing, model inference, and downstream calls. Browse 27,000+ ready-made workflows →
Turing test: Turing's 1950 proposal that a machine be judged intelligent if a human cannot distinguish its written answers from a person's. Modern chatbots pass casual versions of it while still failing basic reasoning, which is mostly a comment on the test.

U

Unattended automation: Automation that runs independently on servers or virtual machines without a person present. It requires scheduling, credential management, monitoring, and recovery controls. Browse 27,000+ ready-made workflows →
Unbounded consumption: Allowing AI workloads to consume excessive tokens, compute, tool calls, time, or money. Limits, quotas, timeouts, budgets, and anomaly detection reduce the risk.
Underfitting: A model too simple to capture the real pattern — poor on both training and new data. The opposite failure to overfitting.
Unsupervised learning: Finding structure in unlabelled data — clustering, dimensionality reduction. No ground truth means evaluation is a judgement call.
User prompt: The request or content supplied by the end user. User prompts are untrusted input and may be incomplete, ambiguous, adversarial, or contain sensitive information. Browse the prompt library →

V

Vector database: A database (Pinecone, Qdrant, pgvector) that stores embeddings and finds semantically similar content — the storage layer behind RAG.
Validation set: Data held back to tune decisions during development. Kept separate from the test set, which should be touched once.
Vector: An ordered list of numbers used to represent an item in a multidimensional space. In semantic retrieval, vector proximity approximates similarity in meaning. Explore AI tools →
Vector and embedding weakness: Security or reliability problems in embedding and retrieval systems, including unauthorized retrieval, poisoned documents, cross-tenant leakage, weak filters, and misleading similarity matches. Explore AI tools →
Vector search: Finding records whose embeddings are nearest to a query embedding. Retrieval by meaning rather than by matching words. Explore AI tools →
Vendor lock-in: The cost or difficulty of moving away from a platform because workflows depend on proprietary connectors, formats, features, data, or pricing. Portable schemas and modular architecture reduce lock-in.
Verifier: A separate check confirming an agent's output before it is accepted or acted on. Explore the AI agents directory →
Vibe coding: Building software by describing intent to a model and iterating on what it produces, rather than writing most lines yourself.

W

Workflow automation: Connecting apps and services so multi-step business processes run without manual work. Platforms like n8n, Zapier, and Make let you build these flows visually from triggers and actions. Full explanation →
Webhook: An HTTP callback that lets one system notify another the moment something happens, instead of the receiver polling for changes. The backbone of real-time automation. Full explanation →
Workflow template: A pre-built, importable automation you adapt to your accounts and data — hours of configuration compressed into minutes. 27,000+ workflow templates →
Watermarking: Embedding a detectable signal in generated content to mark its origin. Fragile against editing and re-encoding. Explore AI tools →
Weak AI: AI built for one narrow task — the only kind that currently exists. Also called narrow AI.
Webhook verification: The process of confirming a webhook's signature, timestamp, source, and replay window before trusting its payload. A public webhook URL alone is not authentication. Browse 27,000+ ready-made workflows →
Worker: A process that takes jobs from a queue and performs them. Scaling the number of workers increases capacity until another dependency becomes the bottleneck.
Workflow: A defined sequence of steps that transforms an input or event into an outcome. A workflow may be deterministic, AI-assisted, agentic, manual, or a combination of these approaches. Explore the AI agents directory →
Workflow instance: A specific running or completed occurrence of a workflow definition with its own inputs, state, and execution history. One workflow definition can produce thousands of separate instances. Browse 27,000+ ready-made workflows →
Working memory: What an agent is actively holding for the current task, bounded by the context window. Explore the AI agents directory →

Z

Zap: Zapier's name for an automated workflow: one trigger plus one or more actions. Zapier workflow templates →
Zero-shot learning: Asking for a task with no examples, relying on what the model already learned. Browse the prompt library →
Zero-shot prompting: Asking a model to perform a task using instructions without worked examples. It is simple and token-efficient but may be less consistent for specialized formats. Browse the prompt library →

About this AI glossary

What is an AI glossary?: An AI glossary is a reference list defining the vocabulary used across artificial intelligence and automation — model architectures, prompting techniques, agent patterns, and the platform-specific words that surround them. This one covers 483 terms in plain English, with each definition linked to the tools or templates that actually use it.
What AI terms should I learn first?: Start with the eight that everything else is built on: large language model, token, context window, prompt, hallucination, RAG, AI agent, and workflow automation. Almost every other term in this glossary is a refinement of one of those, so learning them first makes the rest read as variations rather than as new concepts.
What is the difference between AI, machine learning and deep learning?: They nest. Artificial intelligence is the broad goal of software doing things that normally need human judgement. Machine learning is the dominant way of achieving it — learning patterns from examples instead of following written rules. Deep learning is a kind of machine learning that uses many-layered neural networks, and it is what made modern language and vision models work.
How often is this AI glossary updated?: Continuously. The agent, MCP and skills vocabulary in particular is only months old and still settling, so terms are added as they stabilise rather than on a fixed schedule. Definitions are written in-house rather than aggregated, so each one can be corrected when the usage shifts.
Is this AI terminology guide free?: Yes. All 483 terms are on one page with no signup, no account and no paywall, and the eight most-searched terms have their own full explanations.

Put the vocabulary to work

Skip the theory — start from 27,000+ ready-made workflows, or convert one between platforms.

Browse workflows Workflow converter