GLOSSARY

AI model glossary

Terms you meet in this catalog and in talks with a vendor. Short, no formulas: what it means and why it matters to you.

50 termsUpdated 22 Sep 2026Open the full catalog with filters

Approximate nearest neighbour search: Fast approximate search across millions of numeric fingerprints. Checking every item exactly would be far too slow, so the system deliberately trades a little accuracy for speed. In practice this means the right document occasionally misses the results, so search quality is something you tune rather than assume.
ASR (speech recognition): Turning audio into text. Everything involving calls, meetings and voice messages starts here. Judge it on your own audio quality and vocabulary: a studio recording is recognised far better than a phone line, and product names and surnames are what models get wrong most often.
Batch: A group of requests processed together rather than one at a time. Hardware is used far more efficiently that way, so running a thousand documents overnight costs less than the same thousand one by one during the day. The trade-off is that an individual answer waits for its group, so live conversation uses small batches.
Benchmark: A standard set of tasks used to compare models with each other. It is useful for filtering out clearly weak options and useless as a promise of results for you: your documents, language and tasks were not in it. Decisions are made on your own measurements with your own examples, with public tables used only for a first shortlist.
Checkpoint: A saved state of a model at a point in time: a specific version of the files you can run. Different checkpoints of the same model behave differently, so a contract should name the exact version rather than the family. Otherwise an update to what looks like the same model quietly changes the behaviour of a live service.
Context: How much text a model holds in view at once: your question, attached documents and earlier messages. Anything outside the context window simply does not exist for the model. That is why a three-hundred-page contract is not handed over whole; the relevant clauses are found first and only those are passed in.
Data licence: A separate question from the weights: what the model was trained on and what rights that implies for what it produces. Training data composition is often undisclosed and disputes about it continue. The practical takeaway is one: the more public and visible the material, the more attention its provenance deserves. This is general; a lawyer should review your case.
Dataset: The collection of data a model was trained on, or the set you use to test it. On your side that is usually an export of chat history, a document archive or a folder of camera frames. Assembling an honest test set from real, including awkward, cases usually takes longer than setting up the model, and it decides whether you can trust the result.
Deepfake: Video or audio where a person face or voice is replaced with a synthetic one. Businesses meet it from two sides: as a temptation to make a clip with a famous face, and as fraud where scammers use an executive voice to request a transfer. The first is unacceptable without the person permission, the second calls for confirming payments over a second channel.
Detection: Finding objects in an image and marking where exactly they are. Warehouse counters, restricted zone monitoring and safety equipment checks all rest on it. It runs fast and on site, with no video leaving the premises. Quality is decided less by the model than by cameras, lighting and a clean lens.
Diarization: Splitting a recording by speaker: who said each line. Without it a call transcript is one undivided block, and you can neither assess an agent nor check whether a required script was read. For meetings with several people the same step produces minutes that show who proposed what.
Distillation: Training a small model on the answers of a large one to get nearly the same quality for less money. The small model inherits behaviour on a specific range of tasks, not the breadth of the original. For a business it is a way to relieve expensive hardware: the small model handles the bulk flow and hard cases go to the big one.
Embedding: A numeric fingerprint of the meaning of a text or an image. Texts with similar meaning get similar fingerprints, so search starts matching substance rather than exact words. A customer writes about a broken lock and the guide on repairing fittings comes up, even though the two share almost no wording.
Fine-tuning: Further training of a ready model on your own examples so that it picks up your answer format, terminology or style. It needs a collected set of examples and has to be repeated after any significant change. Facts like prices and deadlines are not taught this way: they live in documents and are retrieved, otherwise every price change means retraining.
Fine-tuning versus prompting: A fork that appears in almost every project: train the model on your examples, or simply write a good instruction and supply the documents. Start with the second almost every time: it is cheaper, faster and fixable in minutes. Fine-tuning earns its place when you need a stable format and voice, not when facts are missing.
GGUF: A popular file format for distributing compressed models that run on ordinary hardware. A single file holds the model and its metadata, so it can be downloaded and started straight away. If a contractor says they will use a GGUF build, that usually means running it on your own machine or a modest server instead of renting an expensive graphics card.
Guard model: A small fast model that writes nothing and only checks: the incoming message for attempts to extract what it should not, and the outgoing answer for unacceptable content. It sits beside the main model as a filter on the way in and out. For public channels it is mandatory, because one bad answer costs more than the whole saving.
Hallucination: A confident falsehood: the model invents a clause, a link or a number that never existed. This is not a fault but a property: the model picks a plausible continuation of text rather than checking facts. It is not fixed by pleading in the instructions, but by grounding answers in retrieved documents and checking them before a human sees them.
Inference: A finished model doing its everyday job: you send a question or an image and get an answer back. Training happens once and is expensive, while inference happens all day and makes up nearly all of the running cost. When a contractor talks about inference cost, they mean serving your daily flow of requests, not the one-off build.
Labelling: Manual work where people mark the correct answers in the data: where the defect is in a frame, which topic a ticket belongs to, what the invoice total is. Without labels there is nothing to train a narrow model on and no honest way to test a ready one. It is the most underestimated budget line, because it takes time from people who know the subject.
Latency: The delay between a request and an answer, especially before the first words appear. This is what a person feels. In chat a few seconds are tolerable; on a phone call a two-second pause already reads as a dropped line. Latency rather than text quality often decides whether a service is usable in your channel at all.
LLM as judge: A model set to grade another model answers against your criteria: accurate, polite, not invented. It lets you check thousands of answers instead of reading a sample by hand. The judge must be calibrated against human grades on a small sample, otherwise you get automated quality control that is quietly wrong on its own.
LoRA: A cheap way of fine-tuning: instead of reworking the whole model you train a small add-on layer on top of it. The file is small, training is fast, and the add-on can be switched on and off. It is a convenient way to keep several behaviours on one base model, for example a separate answer style for each brand inside a company.
Mixture of experts (MoE): A model built from many parts where only a small share of them switches on for each request. The result is a model with broad knowledge that computes almost as fast as a small one. The catch is that the whole thing still has to sit in memory, so you save on response speed rather than on graphics card requirements.
Multimodality: The ability of one model to handle not only text but also images, audio or video. It helps when a single customer message contains both a photo and a caption. At the same time a narrow model built for one data type is usually more accurate than a general one, so working setups often use both rather than picking one.
OCR: Text recognition on an image: a scan, a photo of a document, a camera frame. Modern models of this class return page structure as well as letters: tables, columns, footnotes. It is the first link in any document workflow and its quality caps everything downstream: you cannot extract correct data from a badly recognised invoice.
Ollama: A simple program that installs and runs models on an ordinary computer: it downloads the file, starts a local service, and applications talk to it. It suits trials, demos and a small team. For a real flow of customer requests it is usually replaced with an industrial serving setup built for many users at once.
Open weights: The files of a finished model that the developer published for anyone to download and run on their own hardware. This is not the same as fully open source: the training data is usually not released. For a business, open weights mean your data stays on your side and the service does not stop if a vendor raises prices or closes access.
Overfitting: The model has memorised your training examples and performs beautifully on them while getting lost on new cases. From the outside this looks like a brilliant demo and weak results in real work. The defence is simple: set part of the examples aside in advance, never show them during training, and measure only on those.
Parameters (B): Model size measured in billions of internal numbers: 7B means seven billion, 70B means seventy. Bigger usually means better answers, but also costlier hardware and slower responses. For a narrow job such as sorting incoming tickets, a small model often does just as well and costs far less to run.
Pipeline: The chain of steps a request passes through: recognise, retrieve, answer, verify. Almost every working deployment is built this way rather than as one model doing everything. The practical value is that each link can be measured and replaced on its own, and when quality drops you can see which step broke.
Prompt: The text you give a model: the question, the task, any attached data and an example of the answer you want. Prompt quality affects the result more than choosing between two comparable models. In practice, half of failed rollouts are fixed not by swapping the model but by writing a clear task with examples of a correct answer.
Prompt injection: A trick where an instruction is hidden inside an incoming email, document or web page, and the model follows it instead of yours. The risk appears wherever an assistant reads outside text and has access to data or actions. The defence is an input filter, restricted permissions, and a rule that retrieved text counts as data rather than a command.
Quantization: Compressing a model by storing its internal numbers more coarsely so it needs less memory and runs faster. Quality drops a little, usually unnoticeably in chat and more visibly in precise calculations. In practice this is the difference between needing an expensive server card and running the same model on an ordinary desktop machine in the office.
RAG (retrieval-augmented generation): An approach where the model first searches your documents for relevant pieces and then answers based on them. That keeps answers tied to your policies, prices and deadlines rather than general knowledge from the internet. It is the main way to build a company assistant without retraining anything, and updating it is just replacing a document.
Reranker: A model that reorders retrieved fragments while reading the full question properly. Fast search returns a dozen similar pieces, but the right one is often not first. A reranker costs very little compared with a large model and noticeably improves support answers, because the model receives the correct paragraph rather than the one next to it.
Segmentation: Outlining an object by its contour rather than with a rectangle. That matters wherever you need to measure rather than merely notice: the area of a chip, the width of a gap, the shape of a stain. In retail the same technique removes the background from a product photo so catalogue cards look uniformly tidy.
System prompt: A standing instruction attached to every request that sets the frame: who the assistant is, how to answer, what not to do, when to hand over to a human. Customers never see it, but it defines tone and boundaries. Changing one line in it changes the behaviour of the whole bot, so its versions deserve the same care as code.
Temperature: A setting that controls how varied the answers are. Low temperature gives predictable, dry text; high temperature makes it livelier but raises the risk of invention. For pulling fields out of invoices it is set near zero; for advertising drafts it is raised. It is the cheapest knob for adjusting behaviour without changing the model.
Throughput: How many requests the system digests per unit of time. This is what the budget feels. Latency and throughput pull against each other: settings that speed up an individual answer usually reduce the total number of requests served, so the choice is made per scenario rather than in general.
Token: The chunk of text a model counts in: roughly a word or part of a word. Tokens measure the size of your request and of the answer, and cloud services bill by them. The practical meaning is simple: a long instruction attached to every request is paid for again each time and slows the answer down.
TTS (speech synthesis): Turning text into a spoken voice. It is used in telephony, video voice-over and audio notifications. Besides how clean the voice sounds, the delay before speech starts matters: in conversation a pause longer than a couple of seconds feels like a dropped call. Cloning a specific person voice requires their permission.
Upscaling: Increasing the resolution of an image or video while filling in detail. It helps stretch small product photos up to a marketplace requirement or bring old footage to a decent state. Remember that the model invents something plausible rather than restoring what was lost: small text on a label can come out fabricated after upscaling.
Vector database: Storage for those numeric fingerprints along with links to the source documents. A regular database matches exact values, this one matches closeness of meaning. For a business it is a technical detail with one practical consequence: when a policy changes you must also recompute its fingerprints, or search keeps returning the old version.
Video memory (VRAM): Memory on the graphics card, where a model must fit entirely to run fast. This is the main constraint when choosing: not processor speed or disk size, but how many gigabytes the card has. If a model does not fit, it either will not start or answers many times slower, which kills live conversation and support scenarios.
Vision-language model (VLM): A model that sees an image and understands text at the same time: you can show it a photo and ask a question about it. This is what powers product descriptions from a shot, receipt analysis and explaining what a camera noticed. Unlike text recognition, it understands the content of a scene rather than only the letters in the picture.
vLLM: An industrial way to serve models on a server: it handles many simultaneous requests, uses graphics card memory efficiently and streams answers back. If you expect a flow of requests rather than one employee experimenting, this is what a contractor will propose. The difference from a desktop app shows not in answer quality but in how many people one machine serves.
Watermark: A hidden or visible mark showing that material was generated. Some models add one automatically and some do not, and removing it is not hard, so it is not a defence you can rely on. A company is better served by its own rule: label synthetic content in your materials and keep the originals.
Weights: The numbers inside a model that hold everything it learned. Without weights a model is an empty shell: the architecture is known but there is nothing to answer with. Weights range from hundreds of megabytes to hundreds of gigabytes, and they are what you download, store and update when you say a model has been installed on your server.
Weights licence: The document stating what you may do with a downloaded model: whether it can go into a commercial product, whether there are limits by number of users or by industry. Terms vary a lot between families, and freely downloadable files do not equal permission for any use. This is a general description; a lawyer should review your case.

Need a model for your task?

An open model can run on your own server: data stays in-house, there is no per-request fee, and the model can be fine-tuned on your documents.

SelectThe model and size for your task and hardware budget
DeployOn your server or in a closed network, with an API
Fine-tuneOn your data, or connect a knowledge base
IntegrateInto your CRM, ERP, bot, website or team chat

Discuss deployment