AI glossary: generative AI, decisions and evaluation

A neural network trained on very large amounts of text to predict the next token. That single skill lets it write, summarise, translate, classify and answer questions; most chat assistants and generative AI products run on one.

See also: Autoregressive generation Generative AI

A large model trained on broad data and then adapted to many tasks by prompting, retrieval or fine-tuning, instead of being built for a single one. LLMs and many vision and multimodal models are foundation models.

A model that accepts or produces more than one kind of data (text, images, audio or video) within a single system. Reading a scanned form and answering questions about it is one example.

A model whose trained weights are published, so it can run on your own infrastructure. Licences differ widely: open weights do not automatically mean open training data, open source code or unrestricted commercial use.

In practice Often the deciding option when data has to stay inside your own perimeter.

See also: Edge & On-Premise AI

The prompt is the input a model receives for a task. A system prompt is a standing instruction set by the application, covering things like role, rules and output format, and it applies to every request.

In practice Instructions in a prompt are guidance, not enforcement. Anything that must never happen belongs in application code.

See also: Prompt injection

The maximum amount of text, measured in tokens, that a model can take into account in one request. Instructions, documents and conversation history all count toward it.

In practice A larger window is not the same as better use of it. Long inputs full of irrelevant detail can still degrade the answer.

Retrieving relevant passages from your own documents at query time and giving them to the model as context, so answers draw on current, company-specific information without retraining the model.

See also: Fine-tuning vs. RAG Embedding

A list of numbers that represents the meaning of a piece of text or an image, so that similar content ends up close together. Embeddings power semantic search and the retrieval step in RAG.

A database built to store embeddings and find the nearest ones quickly. It is the usual search index behind a RAG system.

Continuing to train an existing model on your own examples so that it adopts a task, style or output format more reliably. It changes behaviour; it is a poor way to keep a model current with changing facts, which is what retrieval is for.

See also: Fine-tuning vs. RAG

Tying a generated answer to specific source material and pointing to it, so a reader can check where each claim comes from.

In practice A citation shows where an answer claims to come from. Whether the source actually says that still has to be checked.

A model producing a structured request to run a defined function, such as searching a database, creating a ticket or calling an API. The application runs the function and passes the result back to the model.

See also: AI agent

A system in which a model plans and carries out several steps, calling tools and acting on their results, to complete a task rather than answer a single prompt. Once it changes state in real systems, it needs the permissions, evaluation and rollback of production software.

See also: Agentic AI & Workflow Automation Approval gate

Content hidden in an input, such as an email, a web page or a document, that tries to override the model’s instructions, for example to leak data or trigger an action.

In practice The main defence is architectural: narrow permissions and approval gates, not a better-worded prompt.

See also: Approval gate

Checks around a model that filter or block inputs and outputs, for example for off-limits topics, personal data, format or policy. They reduce risk but do not replace permissions or evaluation.

A language model trained to work through intermediate steps before giving its final answer. It tends to do better on multi-step problems, at the cost of more tokens, more latency and a higher price per answer.

Running a trained model to produce an output, as opposed to training it. Inference cost and latency are what you pay for in production, usually per request or per token.

See also: Pricing AI per use

Two ways to make a model smaller and cheaper to run. Quantization stores its weights at lower numerical precision; distillation trains a smaller model to imitate a larger one. Both trade some quality for speed and cost, and that trade has to be measured rather than assumed.

See also: Edge & On-Premise AI

The date after which a model’s training data ends. It knows nothing later unless that information is supplied at query time, for example through retrieval.

See also: Retrieval-augmented generation

How most large language models produce text: one token at a time, each chosen based on the input and every token generated so far. The output is built sequentially rather than returned in one step.

In practice Autoregressive does not automatically mean random. Whether the output varies depends on how each next token is chosen.

See also: Greedy decoding Sampling and temperature

The unit a language model reads and writes: a word, part of a word or a punctuation mark. Pricing, context limits and generation speed are all measured in tokens.

Choosing the single most likely next token at every step instead of sampling. It makes generation far more repeatable, though not necessarily correct, and differences in serving infrastructure can still cause occasional variation.

See also: Determinism vs. reproducibility

Choosing the next token at random, in proportion to its probability. Temperature controls how widely that choice spreads: low values concentrate on the likeliest tokens, higher values spread it out.

In practice Useful for varied writing; usually unwanted for decisions that should come out the same way every time.

A sampling method that restricts the next-token choice to the k most likely candidates before sampling among them. It shapes how text is generated; it is not a list of alternative answers to a business question.

A model response in a defined, machine-readable format, for example JSON with fixed fields, instead of free text. Language models can produce structured outputs; decision models return them natively.

In practice A valid structure says nothing about whether the values inside it are right.

See also: Schema-constrained decoding Jev by TypeSafe

Restricting a language model’s token choices during generation so that the output is guaranteed to match a given schema, such as a JSON schema or a fixed list of labels.

A model built to return a bounded judgment instead of generated text: a choice among defined options, a score on a rubric or a yes/no probability. TypeSafe calls its version a System One model, and Jev is the first public example.

See also: Jev by TypeSafe

A fluent, confident output that is not supported by the input or by fact.

In practice Guaranteed schema matching removes malformed answers, not wrong ones. A “zero hallucinations” claim should state which of the two it means.

A probability for every allowed answer, adding up to one. For a choice between four teams, the model returns four numbers rather than a single label.

Picking the option with the largest probability. Given the probabilities, the choice is fixed: an option at 85 percent is selected every time, not 85 percent of the time.

How well stated probabilities match observed frequencies. Among all the cases a well-calibrated system rates at 80 percent, about 80 percent turn out to be right.

In practice A rule such as “route automatically above 90 percent” is only meaningful if calibration has been measured on your own cases.

See also: What a useful evaluation measures Decision threshold

In everyday usage, how sure a system is. In specific products it can mean something narrower: Jev’s confidence field, for example, is a statistic derived from the shape of the probability distribution, not the probability that the answer is correct.

In practice Check what a vendor’s confidence number is computed from before using it in a business rule.

The probability or score at which a workflow acts automatically instead of sending a case to review.

In practice A business policy set from measured error costs rather than a number copied from a demo. A wrong queue assignment and a wrong payment approval deserve different thresholds.

See also: Human review

The confirmed correct answer for a case, established from the real outcome or by the accountable experts.

In practice Labels generated by another model are a reference, not ground truth.

A fixed collection of representative cases with ground-truth answers, used to score every candidate model and every update the same way. It usually outlives the models it is used to compare.

See also: AI Evaluation & Testing

The share of cases a system gets right against ground truth. Useful but incomplete: it counts every error the same, whatever the mistake would cost.

Whether the same case, run again with unchanged inputs on the same model version, receives the same or a similar judgment.

In practice A system can be perfectly consistent and consistently wrong. Test consistency and accuracy separately.

A deterministic process gives identical output for identical input by construction. Reproducibility is the observed property that reruns match. Many AI systems are reproducible enough in practice without being strictly deterministic. Neither property says anything about correctness.

How often two models give the same answer on the same cases.

In practice High agreement shows that the models behave alike. It does not show that either of them is right.

See also: Understanding and verifying is the new bottleneck

Using a language model to grade other outputs against a rubric. Fast and cheap at scale, but its grades inherit the judge model’s own errors and biases, so they need periodic checks against human ground truth.

A standard test set and metric used to compare models. Useful for tracking the field; weak evidence for a specific workflow unless its cases resemble your own.

See also: Five questions before you trust a number

A regression is a drop in quality after a change, such as a new model version, prompt or tool. Drift is gradual change in the inputs or in the model’s behaviour over time. Both are caught by rerunning a fixed evaluation set.

Fixing a request to an exact model version instead of a moving alias such as “latest”. An alias can switch the model behind an unchanged request.

In practice Record the pinned version with every evaluation result, and test a new version before it reaches production.

A training method that tunes a language model toward the responses people prefer, using human preference ratings to train a reward model. It shaped most of today’s chat assistants.

TypeSafe’s name for the training approach behind Jev. Its stated objective is structured decisions with probabilities that reflect how often those decisions are correct; the public materials do not describe in detail how it works.

See also: Calibration Jev by TypeSafe

A step where a person checks or approves a model’s output before the workflow acts on it. It can apply to every case, or only to cases below a threshold or flagged by a rule.

A rule in the application that blocks an irreversible action (sending, paying, deleting) until an authorised person confirms it. It belongs in code, not in an instruction to the model.

See also: Agentic AI & Workflow Automation

What a system does when the model times out, returns nothing or meets an input it does not support. A defined fallback, usually human review, keeps a missing answer from turning into a wrong action.