Glossary of AI, LLMs and data

A statistical comparison of two or more techniques. Technique A is usually the existing one, B the new one. The test shows which performs better and whether the gap is statistically significant. It usually looks at a single metric, but can compare several.

Velista illustration Compare the current pricing rule (A) with a rule recommended by a model (B), across comparable stores.

Source Google, Machine Learning Glossary

See also in the glossary: Cannibalisation

Regulation (EU) 2024/1689, presented by the European Commission as the first comprehensive legal framework on AI. It sets rules across four risk levels: prohibited practices (unacceptable risk), high-risk systems, transparency risk (for example telling users they are talking to a machine, making generated content identifiable) and minimal risk, with no specific rule. It also provides rules for general-purpose AI models.

Timeline according to the Commission (page updated on 3 August 2026): entered into force on 1 August 2024, applicable on 2 August 2026 with exceptions. The rules on general-purpose models have applied since August 2025. The rules for the high-risk uses listed in Annex III are postponed to 2 December 2027, those for systems embedded in products to 2 August 2028, by Regulation (EU) 2026/1744 (Digital Omnibus on AI).

Source Commission européenne, AI Act

Software that can reason about a user's requests to plan and carry out actions on their behalf. Anthropic contrasts agents, where the LLM directs its own process and its use of tools, with workflows, where the path is set by code (see Workflow). Its shortest formulation: LLMs using tools autonomously in a loop. In reinforcement learning the word means something else: the entity that applies a policy to maximise the expected return.

Sources Google, Generative AI glossary ; Anthropic, Building effective agents ; Anthropic, Effective context engineering

See also: Workflow and agentic workflow

A set of features and rules in a piece of software that lets other programs interact with it automatically, rather than through its user interface.

Velista illustration A SaaS vendor exposes an API so its customers can pull its recommendations into their own systems.

Source MDN Web Docs, Glossary: API (in French)

A non-human program or model that can solve sophisticated tasks, for example translating a text or identifying diseases on medical images. Machine learning is formally a subfield of AI.

Source Google, ML glossary (fundamentals)

Google distinguishes two forms, in the sense of ethics and fairness. The first: stereotypes, prejudice or favouritism towards some things, people or groups, which can affect the collection and interpretation of data, the design of a system and how people interact with it (automation bias, confirmation bias and so on). The second: systematic error introduced by a sampling or reporting procedure (selection bias, coverage bias, non-response bias and so on).

Not to be confused with the bias term (intercept) of a model, nor with prediction bias.

Source Google, ML glossary (fundamentals)

See also in the glossary: Survivorship bias

NIST definition (SP 1500-1): large datasets, characterised mainly by volume, variety, velocity and/or variability, that require a scalable architecture for efficient storage, manipulation and analysis.

Source NIST, Big Data

A small model designed to run on devices with limited computing resources: mobile phones, tablets, embedded systems. Google cites Gemini Nano as a model designed for on-device use.

Sources Google, Machine Learning Glossary ; Google, Generative AI glossary

A change in the relationship between the features and the label. The model learns this relationship on its training set; if it stops reflecting the real world, its predictions degrade over time. Google's example: the notion of a “fuel-efficient car” changes, so a model labelled in 1994 would almost certainly be wrong in 2024. Google recommends retraining faster than the pace of the drift. See also Nonstationarity.

Source Google, Machine Learning Glossary

See also: Nonstationarity

See also in the glossary: Health score, Data freshness

According to Google, the number of tokens a model can process for a given prompt: the larger the window, the more information the model has to produce coherent responses. According to Anthropic, the amount of text a language model can look back on and use to generate new text. A larger window allows longer and more complex prompts; a smaller one can limit coherence over long exchanges.

Sources Google, Generative AI glossary ; Anthropic, glossary

According to Anthropic, a set of strategies for selecting and maintaining the optimal set of tokens (information) during LLM inference, including everything that lands in the context outside the prompts. Anthropic presents it as the natural successor to prompt engineering, whose main focus is writing effective prompts, in particular system prompts. With agents working over multiple turns, you have to manage the whole state of the context: system instructions, tools, MCP, external data, message history.

Source Anthropic, Effective context engineering for AI agents

In a clinical trial, according to the National Cancer Institute, the group that does not receive the new treatment being studied; it is compared with the group that does, to find out whether the treatment works. The EU terminology database describes it as the reference against which a comparison is made in an experiment, typically used to measure the effect of a treatment by comparing the control group with the experimental group.

Velista illustration Apply price recommendations to one batch of stores and keep a batch of comparable stores on the usual practice.

Sources National Cancer Institute, control group ; EU terminology database (IATE), via Gaois

See also in the glossary: Cannibalisation

A framework attributed to Joseph Juran that splits the costs of quality into four categories: prevention, appraisal, internal failure and external failure. The first two are costs of conformance, the other two costs of non-conformance.

Velista reading, not from a source Applied to AI: prevention is data quality, instructions and guardrails; appraisal is evals; internal failure is rejected or reworked responses; external failure is the error that reaches the customer.

Sources iSixSigma, Cost of Quality ; The Cost of Quality: A Primer

Also in the glossary: Cost of quality

According to Gartner, the specification of decision rights and an accountability framework to ensure appropriate behaviour in the valuation, creation, consumption and control of data and analytics. According to DAMA-DMBOK2, the exercise of authority and control (planning, monitoring, enforcement) over the management of data assets.

Sources Gartner, Data Governance ; Government of Canada (GCwiki), citing DAMA-DMBOK2

According to the UK government, good-quality data is fit for its intended use. It recommends six dimensions, defined by DAMA UK: accuracy, completeness, uniqueness, consistency, timeliness and validity.

Velista illustration The same item listed under two references (uniqueness); a competitor price collected ten days ago (timeliness).

Sources GOV.UK, Meet the data quality dimensions ; GOV.UK, What is data quality?

Also in the glossary: Data quality

A collection of raw data, most often (but not only) organised in a spreadsheet or a CSV file. An example generally corresponds to one row of the set: features and, possibly, a label.

Source Google, ML glossary (fundamentals)

According to IBM, a supply chain process that relies on historical data for demand planning and anticipates future customer demand. Methods range from trend projection to sophisticated AI-driven predictive analytics models. AWS describes it as a field of predictive analytics that seeks to forecast customer demand.

Velista illustration Forecast next week's sales of a promoted item to adjust supply and price.

Sources IBM, What is demand forecasting? ; AWS, Demand Forecasting (whitepaper)

See also in the glossary: Repricing

Reducing a large model (the teacher) to a smaller model (the student) that imitates its predictions as faithfully as possible. Advantages of the student: faster inference, lower memory and energy use. Trade-off: its predictions are generally worse than the teacher's.

Source Google, Generative AI glossary

An array of floating-point numbers taken from a hidden layer of a neural network and describing the inputs to that layer; most often, the vector learned by an embedding layer. These are not random numbers: similar items have similar sets of numbers. The values change with each retraining, even on the same data.

Source Google, Machine Learning Glossary

Evaluation is the process of measuring the quality of a model or comparing several models. A supervised model is judged on a validation set and a test set. “Evals” most often refers to LLM evaluations: metrics and benchmarks that help spot what to improve, compare LLMs and check that they are safe. Three modes exist: automatic (a program compares the output with a reference answer, or an “autorater” judges), human, and mixed (an autorater trained on human judgements).

Source Google, Generative AI glossary

According to Palantir, which says it invented the role, an engineer placed directly with customers, often on site, to quickly understand their most important problems, then design and implement data-driven solutions. In its recent job postings, Palantir describes the mission as deploying software and AI in the high-stakes operations of large institutions, with “a radical commitment to the outcome”. This definition comes from the vendor that originated the term; other companies use the title with varying scopes, not covered here.

Source Palantir, Forward Deployed Software Engineer job posting

An additional, task-specific training pass applied to a pre-trained model to refine its parameters for a particular use, typically with hundreds or thousands of targeted examples. Google distinguishes changing all the parameters (full fine-tuning), changing only some of them, adding layers, and parameter-efficient techniques such as LoRA. Google compares it with distillation and prompt-based learning.

Source Google, Generative AI glossary

A very large model pre-trained on an enormous and varied training set. It responds well to a wide range of requests and can serve as a base model for fine-tuning or other customisation.

Source Google, Generative AI glossary

An emerging field with no formal definition. Most experts agree that its models create content that is at once complex, coherent and original: text, images, audio, music, video. Older technologies (LSTMs, RNNs) can also generate coherent content; opinions differ on whether they belong to generative AI. Google contrasts it with predictive ML.

Source Google, Generative AI glossary

In NVIDIA NeMo Guardrails vocabulary, rules (“rails”) applied at different moments. An input rail acts before the request reaches the model; it can reject or alter it, for example by masking sensitive data. An output rail acts after generation and before the response goes back to the user; it can reject or alter the output. Other rails govern the dialogue, document retrieval and the execution of the tools called by the LLM. The term is not standardised: this definition describes one vendor's framework.

Source NVIDIA, NeMo Guardrails Terminology

The production, by a generative AI model, of an output that looks plausible but is factually incorrect, presented as a claim about the real world. Google also gives the synonym “confabulation”, which it considers technically more accurate, while noting that “hallucination” caught on first.

Sources Google, Generative AI glossary ; Google, Machine Learning Glossary

A loosely defined expression, according to Google. It can refer to a policy of looking critically at generative AI outputs, or to a strategy that ensures people help shape, evaluate and refine a model's behaviour. Google's example: an AI generates code that engineers review.

Source Google, Generative AI glossary

In classic ML, the process of making predictions by applying a trained model to unlabelled examples. For an LLM, the process of using a trained model to generate a response to a prompt. The word has a somewhat different meaning in statistics.

Source Google, ML glossary (fundamentals)

The ability to explain or present a model's reasoning in terms a human can understand. According to Google, most linear regression models and decision forests are highly interpretable. Conversely, a “black box” model is one whose reasoning is impossible or hard for humans to understand: this is the case for most deep models and LLMs.

Sources Google, ML glossary (fundamentals) ; Google, Machine Learning Glossary

A feature is an input to a model; an example includes one or more features. The label is, in supervised learning, the “answer” or result of an example: “spam” or “not spam” in a spam detection dataset, the amount of rain that fell in a weather dataset. A labelled example pairs features with a label; an unlabelled example is used for inference.

Source Google, ML glossary (fundamentals)

According to Google, at a minimum a language model with a very large number of parameters; more informally, any language model based on the Transformer architecture. All these models are autoregressive: they predict the next token from the tokens already produced. Anthropic similarly describes autoregressive language models as pre-trained to predict the next word from the document's context.

Sources Google, Generative AI glossary ; Anthropic, glossary

A program or system that trains a model from input data. The trained model can make useful predictions on new data drawn from the same distribution as the training data. The term also refers to the corresponding field of study. ML is formally a subfield of AI, even though some organisations now use the two terms interchangeably.

Source Google, ML glossary (fundamentals)

According to Anthropic, an open protocol that standardises how applications provide context to LLMs. It lets AI systems access external resources in a standardised way. The “MCP connector” is a feature of Anthropic's API that lets you connect to MCP servers without building an MCP client.

Source Anthropic, glossary

According to Google Cloud, an ML engineering culture and practice that aims to unify the development (Dev) and operation (Ops) of ML systems. Practising MLOps means advocating automation and monitoring at every step of building an ML system: integration, testing, release, deployment and infrastructure management. Google's document covers continuous integration (CI), continuous delivery (CD) and continuous training (CT).

Source Google Cloud, MLOps: continuous delivery and automation pipelines

A mathematical construct that processes input data and returns an output; in other words, the set of parameters and structure a system needs to make predictions. A linear regression model consists of weights and a bias; a neural network, of hidden layers of neurons with their weights; a decision tree, of its shape, its conditions and its leaves. A model can be saved, restored or copied.

Source Google, ML glossary (fundamentals)

A neural network is a model containing at least one hidden layer. A deep model (deep neural network) contains more than one. Some networks can imitate extremely complex nonlinear relationships between the features and the label. Google notes that linear models are generally easier to train and more interpretable than deep models, which in turn learn complex relationships between features.

Source Google, ML glossary (fundamentals)

The characteristic of a feature whose values change along one or more dimensions, most often time. Google's examples: a store's swimsuit sales vary with the season; a fruit harvest is zero for most of the year, then high for a short period. Its opposite is stationarity. Google points out that very few features are stationary in the real world.

Source Google, ML glossary (fundamentals)

According to the Open Source Initiative (Open Source AI Definition 1.0), an AI system is open source if it grants the freedoms to use it, study it, modify it and share it, for any purpose, with or without modification. Exercising these freedoms requires access to the preferred form for modifying the system: sufficiently detailed information about the training data for a skilled person to build a substantially equivalent system, the complete training and execution code, and the model parameters (the weights). The OSI states that open source “models” or “weights” must include the information about the data and code used to produce them.

Source Open Source Initiative, Open Source AI Definition 1.0

A model whose weights are widely available. The US NTIA, in its 2024 report, calls these models “open foundation models” and notes that developers can release them with varying degrees of openness: weights, code, training or fine-tuning data and documentation can each be made available through several channels, with varying restrictions.

Not to be confused with open source: the OSI definition also requires information about the training data and code (see Open source (AI)).

Sources NTIA, report on models with widely available weights ; NTIA, background to the report

See also: Open source (AI)

Overfitting is creating a model that fits the training data so closely that it fails to predict correctly on new data. Regularisation and a large, varied training set reduce it. Generalisation is a model's ability to predict correctly on new data it has never seen: it is the opposite of overfitting. A generalisation curve, which plots training loss and validation loss, signals overfitting when the second becomes clearly higher than the first.

Source Google, ML glossary (fundamentals)

Parameters are the weights and biases that a model learns during training. Hyperparameters are the values you supply to the model (or that a tuning service adjusts) from one training run to the next, such as the learning rate. Temperature is a hyperparameter (see Temperature).

Source Google, ML glossary (fundamentals)

See also: Temperature

According to the CNIL, any information relating to an identified or identifiable natural person. The person can be identified directly (surname, first name) or indirectly (telephone number, postal or email address, voice), from a single piece of data or by combining several. The GDPR governs the processing of such data.

Source CNIL, La CNIL en bref (in French)

The ratio between the relative change in the quantity demanded and the relative change in price. Demand is elastic when its change exceeds that of the price, and inelastic in the opposite case.

Sources Office québécois de la langue française, Grand dictionnaire terminologique (in French) ; Éduscol, L'élasticité-prix de la demande (in French)

Also in the glossary: Price elasticity

According to Google, text entered as input to an LLM to condition it to behave in a certain way. A prompt can be short or very long; Google distinguishes the categories question, instruction, example, role and partial input to complete. Prompt engineering is the art of creating prompts that get the desired responses; it depends notably on the pre-training set, the temperature and other decoding parameters. For Anthropic, its main focus is writing effective prompts, in particular system prompts.

Sources Google, Generative AI glossary ; Anthropic, Effective context engineering

According to OWASP, a vulnerability that occurs when user prompts alter the behaviour or output of an LLM in unintended ways. It is possible because LLMs do not separate instructions from external data. The injection is called indirect when the malicious instruction sits in external content the model reads (web page, file).

Sources OWASP, LLM01 Prompt Injection ; OWASP, LLM01 (2023-24 edition)

According to Google, a technique that improves the quality of an LLM's outputs by grounding them in knowledge sources retrieved after the model was trained; the LLM accesses information from trusted knowledge bases or documents. Motivations cited: increasing factual accuracy, giving access to knowledge absent from training, changing the knowledge used, making it possible to cite sources. In Google's example, the backend searches for the relevant data, adds it to the user's request, then asks the LLM to produce a summary from it. For Anthropic, the model is augmented with an external knowledge base or a set of documents passed into the context window.

Sources Google, ML glossary (fundamentals) ; Anthropic, glossary

In Google's vocabulary, an agent chooses actions through a policy, moves from one state of the environment to another and seeks to maximise the expected return. When human raters score a model's responses to improve it, this is called RLHF (see RLHF).

Source Google, Machine Learning Glossary

See also: RLHF

Using feedback from human raters to improve the quality of a model's responses. Google's example: asking users to rate a response with a thumbs up or thumbs down; the system then adjusts its future responses. Anthropic says it trained Claude with RLHF to make it a more useful assistant.

Sources Google, Generative AI glossary ; Anthropic, glossary

According to Anthropic's documentation, a modular capability that extends Claude's functions. Each skill bundles instructions, metadata and, optionally, resources (scripts, templates) that Claude uses automatically when relevant. Unlike a prompt, which is a conversation-level instruction for a one-off task, a skill loads on demand and avoids repeating the same instructions from one conversation to the next. When a skill triggers, Claude reads its SKILL.md file, whose instructions then enter the context window; scripts are executed and only their output enters the context; other files are read only if the task requires it.

Note The term and format described here are those of Anthropic's documentation.

Source Anthropic, Agent Skills

In supervised learning, the model trains on labelled examples (features and the label, that is, the “answer”) and then predicts on examples without a label. A classification model predicts a class; a regression model predicts a number. In unsupervised learning there is no label: the typical example cited by Google is clustering of related examples, after which a human can give each group a meaning.

Sources Google, ML glossary (fundamentals) ; Google, Machine Learning Glossary

According to Google, a hyperparameter that controls the degree of randomness in a model's output: a high temperature gives a more random output, a low one a less random output. The best choice depends on the application. Anthropic similarly describes it as the parameter that controls the randomness of the model's predictions during text generation.

Sources Google, Generative AI glossary ; Anthropic, glossary

According to Anthropic, the text submitted to the model (a sequence of characters) is encoded into a sequence of tokens that the model processes. Larger tokens improve efficiency in pre-training and inference; smaller tokens make it possible to handle rare or previously unseen words.

Source Anthropic, glossary

According to Anthropic's documentation, a mechanism that lets Claude call functions that you define or that Anthropic provides. Claude decides when to call a tool based on the user's request and the tool's description, then returns a structured call that the application executes (client tools) or that Anthropic executes (server tools).

Source Anthropic, Tool use with Claude

Training is the step where an ML system builds a model from data. Pre-training is the initial training of a model on a large dataset; a pre-trained model is then often fine-tuned, distilled or adjusted to follow instructions.

Sources Google, ML glossary (fundamentals) ; Google, Generative AI glossary

According to Google, asking a generative AI model to create software: the prompts describe the software's goal and functions, which the model translates into source code. The code does not always match the intentions, hence iterations. The term was coined to describe a new kind of coding where you give in to the “vibes”. It originally implied a deliberately loose approach, where you might not even examine the generated code; it has come to mean any form of AI-generated coding. Google distinguishes “conversational coding” (iterative dialogue with the model, the original sense of the term) from “specificational coding” (writing and maintaining a description file of the software in human language, which the model or an engineer turns into software).

Source Google, Generative AI glossary

According to AWS documentation, a service that lets you provision a logically isolated section of the AWS cloud, in which you launch resources within a virtual network that you define.

Source AWS, Virtual Private Cloud

According to Anthropic, a workflow is a system where LLMs and tools are orchestrated through predefined code paths, whereas in an agent the LLM dynamically directs its own process and its use of tools. Workflows offer predictability and consistency for well-defined tasks; agents suit cases where flexibility and model-driven decision-making are needed at scale. Anthropic recommends looking for the simplest solution and adding complexity only when necessary, which may mean not building an agentic system. Google defines the agentic workflow as a dynamic process in which an agent autonomously plans and carries out actions to reach a goal, possibly with reasoning, calls to external tools and self-correction.

Sources Anthropic, Building effective agents ; Google, Generative AI glossary

A zero-shot prompt provides no example of the desired response. A one-shot prompt gives one, a few-shot prompt several (synonym: in-context learning). Few-shot generally produces better results than the other two, at the cost of a longer prompt.

Source Google, Generative AI glossary