AI glossary
What is an LLM? What does RAG mean? And how do AI agents, chatbots and classic automation differ? Terms around AI, language models, AI search and automation – explained simply and technically.
Look it up instead of collecting buzzwords
Search directly for a term or a question. Every answer includes a clear definition, practical context and relevant cross-references.
AEO · Answer Engine Optimisation
Preparing content so that search and answer systems can answer specific questions as unambiguously as possible.
A page answers a common customer question directly, completely and with traceable evidence.
AEO overlaps with SEO and GEO; clear information architecture, precise answers and technical readability help.
Agentic AI · Agent-based AI
An approach in which AI systems plan in several steps, act, check results and adjust their next step.
A system researches data, compares options, asks for approval when uncertain and documents the result.
More autonomy increases the potential benefit and at the same time the surface for errors, the costs and the requirements for oversight and observability.
AI Agent · AI that acts towards a goal
An AI system that interprets a goal, selects next steps and uses tools or data sources to do so.
An agent reads an enquiry, checks an order status via an API and drafts a suitable reply.
Agents can support flexible workflows, but need limited permissions, logging and clear approval points.
AI Automation · AI-assisted automation
Automation in which AI processes unstructured content or prepares decisions.
Incoming emails are summarised and categorised before a fixed workflow forwards them.
AI extends automation to less clear-cut tasks, but needs thresholds, exception paths and human oversight.
AI Guardrails · Technical and organisational guardrails
Rules, checks and limits around the inputs, outputs and actions of an AI system.
An application blocks sensitive data, validates outputs and requires approval before external actions.
Guardrails reduce certain risks, but must be implemented in several layers and tested regularly against real failure cases.
AI Overview · AI-generated search summary
A summary generated by a search engine that appears above or within classic search results.
A search engine answers a question directly and links selected sources next to the summary.
AI Overviews change click paths and visibility; clear sources, reliable content and measurable mentions become more important.
AI Search & Generative Search · AI-assisted and generative search
Search that uses AI to understand information, combine it or formulate it as a direct answer.
Instead of only showing links, a search interface combines several sources into one answer.
Businesses need to provide content so clearly and accessibly that search systems can find, understand and attribute it correctly.
AI Visibility · Visibility in AI answers
A measure of whether and how a brand, organisation or website appears in AI-generated answers.
For relevant questions, a company is named correctly and its website is cited as a source.
AI visibility covers mention, context and source reference; it should be measured with documented test questions rather than individual examples.
API · Application Programming Interface
A defined interface through which software systems exchange data or call functions.
A website sends text to an AI model and receives the answer back via its API.
APIs enable controlled integration into existing systems; costs, limits, availability and data transfer must be planned for.
Artificial Intelligence (AI) · Machines performing cognitive tasks
Systems that perform tasks that would normally require human abilities such as understanding, learning or deciding.
An application sorts support requests, recognises their topic and suggests suitable replies.
AI is an umbrella term. For a sensible assessment, it always has to be clarified which concrete task, data basis and oversight are meant.
Bias · Systematic distortion
A systematic skew in data or results that can unfairly disadvantage certain groups, cases or perspectives.
A system rates certain applicant profiles lower because historical data reflects earlier inequalities.
Bias cannot be solved technically alone; data, the definition of the goal, test groups and impacts have to be considered together.
Citation · Source reference
A reference by an AI or search system to the source of a statement.
A generated answer links the specific expert article an item of information comes from.
Citations increase traceability and can bring qualified traffic, but don’t automatically prove a correct summary.
Cloud AI · Externally operated AI service
An AI model operated by a provider and usually made available via a web interface or API.
An application sends approved content to a model service and receives the processing as an API response.
Cloud AI is quick to use and scalable; contracts, data flows, region, costs and provider dependency need to be assessed.
Computer Vision · Machine understanding of images
AI methods that let computers recognise content and patterns in images or videos.
A quality check highlights visible damage on product photos.
Computer vision can support visual inspections, but must be tested with realistic lighting, camera and edge cases.
Context Window · How much a model can consider at once
The maximum amount of information a model can take into account in a single run.
With a long contract, relevant sections have to be selected if the whole document doesn’t fit into the context window.
A large context window helps with long content, but doesn’t replace good selection, structure or checking of the information provided.
Crawler & AI Crawler · Automated retrieval of web pages
Software that retrieves web pages automatically; AI crawlers do so for AI search, training or other AI services.
A bot retrieves publicly accessible pages and processes headings, texts and links.
Technical accessibility, robots rules and clear page structures influence which content a system can process at all.
Data Protection · Protection of personal data
Rules and measures for the lawful, purpose-bound and secure processing of personal data.
Before using an AI service, the purpose, necessary data, provider, storage location, contracts and deletion periods are checked.
AI doesn’t change the principles of data protection. What matters is the specific data flow, not the label of the tool used.
Deep Learning · Learning with multi-layer neural networks
Machine learning with multi-layer neural networks that can process particularly complex patterns.
Deep learning models recognise objects in images or generate natural-sounding speech.
The method enables many modern AI applications, but usually requires a lot of computing power, data and careful testing.
Diffusion Model · Generative image model
A generative model that often creates images by gradually developing a suitable output from noise.
An illustration for an early campaign concept is created from a text description.
Diffusion models are powerful for visual drafts; rights, brand references and factual image content must be checked.
Edge AI · AI close to the device
AI processing directly on or close to a device instead of sending every input to a central cloud service.
A camera detects an event locally and only transmits the result instead of the entire video stream.
Edge AI can reduce latency and data transfer, but is limited by the computing power, energy and maintainability of the devices.
Embedding · Numerical representation of meaning
A sequence of numbers that represents the meaning or properties of content in a way that allows similarities to be calculated.
A question and document sections are compared as embeddings to find passages that match in content.
Embeddings enable semantic search, but need a model suited to the content and protection of sensitive source data.
Entity · A uniquely identifiable thing or concept
A uniquely identifiable person, organisation, product, place or term with its own properties.
JK Concepts is an entity as a company with a founder, location, services and website.
Unique names, consistent details and relationships help search and AI systems assign content to the right entity.
Evaluation, Evals & Benchmarks · Systematic performance testing
Planned tests that assess models or entire AI systems against defined cases and criteria.
A hundred typical and difficult requests are checked for correctness, source reference, cost and response time.
General benchmarks help with comparisons; what is decisive in addition are your own evals with real business and risk cases.
Few-Shot · Task with a few examples
A model receives a few examples from which it should recognise the format or desired behaviour.
The prompt shows three correctly categorised emails before a fourth is classified.
A few good examples can increase consistency, but use up context space and must cover typical special cases.
Fine-Tuning · Targeted model adaptation
Additional training of an existing model for a particular style, domain or type of task.
A model is tuned with reviewed support dialogues to the desired classification and tone.
Fine-tuning can improve recurring behaviour, but is usually less suitable for current knowledge than connected data access.
Foundation Model · General-purpose base model
A large, generally trained model that serves as the basis for many different applications.
One base model is used for text analysis, support drafts and an internal search in different applications.
Businesses usually use an existing foundation model and extend it with prompts, data access or targeted adaptation.
Generative AI · AI that creates new content
AI that generates new content such as text, images, audio, video or code.
A system turns bullet points into a draft product description.
Generative AI can speed up knowledge work, but it does not produce automatically verified facts and needs clear quality controls.
GEO · Generative Engine Optimisation
Optimising content and technical accessibility for generative search and answer systems.
A specialist page defines terms unambiguously, backs up statements, uses structured data and is technically easy to retrieve.
GEO aims to improve the chance of being found and used as a source correctly, but cannot guarantee a mention or placement.
GPU & VRAM · Graphics processor and graphics memory
A GPU is a highly parallel processor; VRAM is its memory for models, data and intermediate values.
A larger local model needs enough VRAM so that its weights can be loaded fully or efficiently.
GPU performance and VRAM determine which models run locally, how fast they respond and what hardware costs arise.
Grounding · Anchoring in reliable information
Linking an AI answer to provided, verifiable information.
An assistant answers a product question only on the basis of the current documentation and names the sections used.
Grounding can make answers more traceable, but does not fully prevent errors from unsuitable sources or misinterpretation.
Hallucination · A plausible-sounding false statement
Factually wrong or invented model output that can nevertheless be phrased convincingly.
A model cites a source that doesn’t exist or invents a product feature.
Hallucinations make sources, validation and human oversight indispensable, especially for consequential statements.
Human in the Loop · Human oversight in the process
A process design in which people check results, approve decisions or take over difficult cases.
The AI prepares a contract summary and a qualified person checks it before it is used further.
Human oversight is especially important where errors have serious consequences, but must be designed with clear information and a real ability to intervene.
Inference · Running a trained model
Using a trained model to compute an output for a new input.
Every generated chat reply is an inference in which the model computes new tokens from the prompt and context.
For live applications, inference costs, speed and reliability make up a large part of operation.
Jailbreak · Attempt to bypass model safeguards
An attempt to make a model ignore its safety rules through specially crafted inputs.
An input disguises a prohibited request as a role-play or splits it into several seemingly harmless steps.
Models can never be secured against new bypass attempts by prompts alone; the application and permissions must limit the damage.
Knowledge Base · Curated collection of information
A maintained collection of information that people or systems can access in a targeted way.
Product documentation, internal guidelines and approved support articles form a knowledge base.
An AI can only answer as reliably about company knowledge as the content, responsibilities and updating of the knowledge base are organised.
Knowledge Graph · Knowledge as entities and relationships
A representation of knowledge as uniquely named things and their relationships to each other.
A graph connects a company with its products, locations, contacts and certifications.
Knowledge graphs make relationships machine-readable and help with search, data integration and unambiguous entity assignment.
Latency · Response delay
The time delay between a request and the start or completion of processing.
An assistant takes two seconds until the first text appears and eight seconds until the answer is complete.
High latency worsens usability and can slow down automated processes; model, network and orchestration all play a part.
LLM · Large Language Model
A large language model that has learned patterns in text and uses them to analyse or generate content.
An LLM summarises long documents or answers questions in natural language.
LLMs are versatile, but they neither automatically know your current company data nor guarantee correct answers.
Local AI / Local LLM · AI run on your own hardware
An AI model that runs on your own or directly controlled hardware instead of an external model service.
A small language model processes internal texts on a company-owned server without an external API call.
Running AI locally can increase data control, but shifts responsibility for hardware, updates, security and availability to the business.
LoRA · Low-Rank Adaptation
A resource-efficient method for adapting a large model through a small number of additional weights.
An open model receives a small specialised adaptation without retraining all the original parameters.
LoRA reduces the effort and memory needed for an adaptation, but still requires good training data and clean evaluation.
Machine Learning (ML) · Learning patterns from data
A branch of AI in which systems learn patterns from example data instead of receiving every step as a fixed rule.
A model learns from previous orders which transactions look unusual.
The quality of the results depends heavily on suitable data, testing and how similar the training cases are to real-world cases.
MCP · Model Context Protocol
An open protocol through which AI applications can connect tools and context sources in a uniform way.
An AI assistant accesses documents or internal tools through standardised MCP interfaces.
MCP can standardise integrations, but replaces neither authentication, permission checks nor secure tool execution.
Mixture of Experts (MoE) · Combination of specialised sub-models
A model architecture that activates only selected specialised parts of a larger model for each input.
A routing mechanism directs different tokens to the expert layers that suit them.
MoE can make high model capacity usable more efficiently; the total parameter count does not equal the computing cost of each request.
Model · AI model
A trained mathematical system that produces predictions, assessments or new content from inputs.
A language model receives a question and computes an answer from it step by step.
Not every model suits every task. Quality, cost, data protection and operation differ considerably.
Model Distillation · Teaching a smaller model from a larger one
A method in which a smaller model learns from the outputs or behaviour of a stronger model.
A compact model is trained for a clearly defined classification task using examples from a large model.
Distillation can reduce cost and latency, but may also transfer errors or biases of the original model.
Model Drift · Declining model fit in operation
A change in model quality because data, language, processes or real-world relationships differ from the time of testing.
A classification becomes less accurate after new product groups and different wording are introduced.
AI systems need ongoing samples, metrics and repeated tests; a good starting value does not guarantee lasting quality.
Multimodal AI · AI across text, images, audio and video
AI that can process several types of media such as text, images, audio or video together.
A model reads a photo of a rating plate, recognises the serial number and explains the matching maintenance instructions.
Multimodal systems open up documents and real working situations, but need clear rules for sensitive content.
Neural Network · Layered computational model
A computational model of connected layers that learns weights and turns inputs into predictions or outputs.
A neural network assesses whether a document is an invoice, a quote or a delivery note.
Neural networks are the technical basis of many language, image and audio models, but remain hard to judge without evaluation.
OCR · Optical Character Recognition
A technique that turns printed or handwritten text from images and scans into machine-readable text.
Invoice data is recognised from a scanned PDF and prepared for accounting.
OCR makes documents searchable and automatable; recognition errors in numbers or tables still need checking.
Open-Source AI · Open AI software
An AI system whose relevant components are available under open licences; the exact scope can vary greatly.
The code for training and operation is open, while the dataset or model weights may have separate conditions.
The term is used inconsistently. Before use, licences, available components and commercial rights need to be checked specifically.
Open-Weight Model · Model with publicly available weights
A model whose trained weights can be downloaded and run on your own infrastructure.
A company runs an available language model on its own server and adjusts its configuration.
Open weights allow more operational control, but don’t automatically mean open training data, free use or low running costs.
Orchestration · Coordinating models and tools
Technical coordination of several models, data sources, tools and process steps.
An application chooses a model, retrieves company data, validates the output and then starts an approval step.
Good orchestration makes complex AI applications testable and operable instead of hiding logic in a single prompt.
Parameters · Model parameters
Learned numerical values in the model that determine how inputs are processed and outputs computed.
During training, a very large number of weights are adjusted so that the model makes better predictions.
More parameters can increase model capacity, but don’t automatically mean better results for a specific business case.
PII · Personally Identifiable Information
Information that can identify a person directly or indirectly.
Name, email address, customer number and a combination of location and contract data can be personal data.
Before AI processing, it must be clarified which personal data is really needed and where it is transferred to.
Prompt · Input to an AI model
An instruction, question or context that an AI model receives as input.
A prompt contains the task, the target tone, relevant facts and the desired output format.
Clear prompts make useful results more likely, but replace neither reliable data nor expert review.
Prompt Engineering · Systematic design of prompts
Planning, testing and improving inputs so that a model performs the desired task more reliably.
An instruction is extended with a role description, examples, limits and a JSON schema and then tested.
Good prompt engineering reduces misunderstandings, but should be seen as part of a tested system rather than a magic formula.
Prompt Injection · Manipulation through injected instructions
An attack in which malicious instructions in inputs or in content that is read are meant to manipulate the AI system.
A document contains hidden instructions intended to make an agent reveal data or call a tool.
As soon as models read external content or use tools, system prompts are not enough; permissions and actions must be limited technically.
Pseudonymisation · Separating data from direct identity
Processing in which direct identifiers are replaced and separate additional information is kept protected.
Before analysis, a customer number is replaced by an internal pseudonym; the mapping is kept separately protected.
Pseudonymisation can reduce risks, but doesn’t automatically make data anonymous and doesn’t remove data protection obligations.
Quantisation · Reducing a model’s numerical precision
Reducing the numerical precision of a model to lower memory requirements and computing effort.
A quantised language model runs on a smaller graphics card or a powerful laptop.
Quantisation can enable local and cheaper applications, but depending on the method it can cost some quality.
RAG · Retrieval-Augmented Generation
A method in which an AI model retrieves suitable information from defined sources before answering.
An internal search finds relevant product documentation; the model formulates an answer from it with source references.
RAG can make answers more current and traceable if retrieval, permissions and source quality are implemented properly.
Retrieval · Information retrieval
Targeted finding of relevant information before a system produces an answer or decision.
From thousands of document sections, the five most likely to match a query are selected.
Poor retrieval gives even a strong model the wrong basis; search quality and permissions are therefore central.
Schema.org · Vocabulary for structured web data
A shared vocabulary that websites use to describe people, organisations, offers, terms and other content.
A glossary uses DefinedTermSet and DefinedTerm to mark up terms and definitions in a machine-readable way.
Schema.org creates a common language for systems; markup alone does not replace helpful visible content, however.
Semantic Search · Search by meaning
Search by meaning and intent rather than only by exactly matching words.
A search for “correct an invoice” also finds a guide titled “Create a credit note”.
Semantic search improves findability with different wording, but should be combinable with filters and classic search.
Speech-to-Text · Automatic speech recognition
Converting spoken language into written text.
A meeting recording is transcribed and then searched for decisions.
The technology makes documentation and search easier, but confidential conversations need suitable data protection and retention rules.
Structured Data · Machine-readable page data
Machine-readable information that describes the content and meaning of a page in defined fields.
A website additionally marks up its organisation, services or defined terms as JSON-LD.
Structured data makes unambiguous interpretation easier, but must match the visible content and be technically correct.
Structured Output · Output in a defined format
Model output that follows a predefined format or data schema.
Instead of free text, the model returns fields such as category, priority and reasoning as validatable JSON.
Structured outputs can be processed further more safely, but still have to be validated technically and checked for content.
Synthetic Data · Artificially generated data
Artificially generated data intended to mimic real structures or cases.
For a prototype, invented customer cases are created with typical but not real personal data.
Synthetic data can make testing easier and reduce data protection risks, but it does not automatically reproduce rare or problematic real cases correctly.
System Prompt · Top-level model instruction
A top-level instruction that defines a model’s role, rules and desired behaviour.
The system prompt requires the model to answer only from approved documents and to state uncertainty explicitly.
System prompts create guardrails, but are no reliable protection against attacks or incorrect outputs.
Temperature · Randomness setting of the output
A setting that, in many models, influences how predictable or varied an output is.
A lower temperature is usually chosen for data extraction, a higher one possibly for brainstorming.
The setting changes how choices are made, but makes an answer neither automatically correct nor fully reproducible.
Text-to-Speech · Speech synthesis
Generating spoken language from written text.
An application reads out instructions in several languages, making them more accessible.
Text-to-speech improves accessibility and usability, but cloned voices raise additional questions of consent and misuse.
Token · A model’s unit of text
A small unit into which a model splits text; a token can be a word, part of a word or a punctuation mark.
A long German compound word can consist of several tokens and therefore use up more context space.
Token counts affect context limits, processing costs and response times of many AI services.
Tokenisation · Splitting text into tokens
Splitting an input into tokens that the model can process.
Before the model processes a text, it is broken down into words, parts of words and punctuation marks.
Different models count the same text differently; this matters for limits, costs and multilingual content.
Tool & Function Calling · Structured tool invocation
A model’s ability to propose a structured call to an external function or piece of software.
The model generates a validated call to retrieve a delivery status from a business system.
Tool calling connects language with real actions; permissions, input validation and confirmation of critical steps remain the job of the application.
Training · Model training
The process in which a model adjusts its internal parameters using many examples.
An image model sees numerous labelled examples and learns to distinguish certain product defects.
Training a model from scratch is too costly for most businesses; existing models, RAG or limited adaptation are often enough.
Training Data · Data a model learns from
Data from which a model learns patterns and relationships during training.
Historical, correctly categorised enquiries serve as examples for later automatic assignment.
Gaps, errors or biases in training data can show up directly in the later results.
Transformer · Transformer architecture
A model architecture that weights the relationships between elements of an input and enables many modern language models.
When processing a paragraph, the model takes into account which words matter for the meaning of other words.
Transformers explain part of the power of modern models, but on their own say nothing about data quality, cost or reliability.
Vector · Numerical feature vector
An ordered list of numbers with which an ML system represents properties of content mathematically.
The vector of a text lies closer to texts with similar content than to unrelated ones.
Vectors are the technical form behind embeddings and similarity search, but are not directly interpretable by humans.
Vector Database · Database for similarity search
A database or search system that stores vectors and efficiently finds similar entries.
Product texts are stored as embeddings so that a question finds matching sections by meaning.
A vector database is often part of RAG, but on its own solves neither content quality, updating nor access rights.
Workflow Automation · Process automation
Automating a predefined process with defined steps and rules.
A validated form triggers filing, a notification and the creation of a task, one after the other.
Classic workflows are often more traceable than autonomous agents and the better choice for stable, rule-based processes.
Zero-Shot · Task without examples
A model is expected to solve a task from an instruction alone, without first receiving suitable examples.
The model classifies an enquiry as sales or support even though the prompt shows no sample cases.
Zero-shot is quick to set up, but can be less consistent with company-specific categories.
Try a more general search term or reset the filters.
From definition to application
Not every business needs a chatbot. What matters is the task, the data, the benefit, the risks and an operation that stays under control.
Discuss an AI project