AI architecture glossary

A small file of extra numbers trained alongside a model, changing its habits without altering the original weights. Cheaper to produce than retraining the model, portable, and small enough to keep several of them, one per task. Usually produced by a method called LoRA.

Taught in Module 1

An AI system given a goal, a set of tools, and permission to work in a loop: plan a step, act, observe the result, and repeat until done, blocked, or handed to a person.

Taught in Module 4

The training that shapes a model’s behaviour towards being helpful and refusing harm. Genuinely strong, genuinely incomplete, which is why production systems add checks outside the model.

Taught in Module 4

The compromise that keeps vector search fast at scale. Rather than comparing a question against every stored passage, the store checks a small set of likely candidates and returns what is probably nearest. Usually that is the same answer; occasionally it is not, and the balance between speed and thoroughness is a setting somebody can change.

Taught in Module 2

Re-running a pipeline over a stretch of past data after a fix, so the correction reaches the history and not only today. Whether one is possible decides how expensive a changed definition turns out to be.

Taught in Module 3

Splitting documents into short passages, ideally one idea each, so retrieval can find precisely rather than approximately.

Taught in Module 2

The model's working memory: the maximum number of tokens it can consider at once, covering instructions, the conversation, supplied documents, and the answer as it grows.

Taught in Module 1

The usual measure of how alike two embeddings are, scored between zero and one by asking how strongly they point in the same direction.

Taught in Module 2

Storage that accepts everything in its raw form and applies structure later. Flexible and cheap; a swamp without curation.

Taught in Module 3

Automated plumbing that copies data between systems and reshapes it on the way, applying definitions as it goes. It fails quietly, which is why monitoring and ownership matter.

Taught in Module 3

Where your data is stored and processed geographically, and whose law applies to it there. A contract question before it is a technology question.

Taught in Module 6

Storage that demands clean, agreed structure before data enters, which makes reporting fast and consistent, and the warehouse rigid.

Taught in Module 3

Describes a process that produces the same result from the same input, every time. Rules, calculations, and ordinary code are deterministic; a language model is not, which is why exact work is given to exact tools and judgement is given to the model.

Taught in Module 4

A long list of numbers capturing what a passage is about, produced by a model trained so that texts with similar meaning land near each other numerically.

Taught in Module 2

European law whose obligations scale with the risk of the use case: a few uses prohibited, high-risk uses carrying oversight and documentation duties, transparency for most of the rest.

Taught in Module 5

Measuring an AI system against a set of real questions with agreed good answers, on a schedule and after every change, so decline is detected rather than reported.

Taught in Module 5

Additional training on your own examples that adjusts the model's weights, changing its habits rather than its knowledge.

Taught in Module 1

The most capable tier of models available at a given moment. The frontier moves; the label follows capability, not any one product.

Taught in Module 1

A fixed set of real questions with agreed good answers, run against a system on a schedule and after every change. The working tool of evaluation, including questions the system must answer and questions it must refuse.

Taught in Module 5

Whether the claims in an answer are supported by the passages the system retrieved. Production systems check it mechanically before delivering.

Taught in Module 4

The controls wrapped around an AI system: permissions, approval gates, budgets, audit logs, and the input and output checks around every model call.

Taught in Module 4

A fluent, confident, wrong answer produced when a model completes a pattern beyond its knowledge.

Taught in Module 2

Autonomy assigned per class of action, with a person approving anything consequential until months of evidence justify loosening, class by class.

Taught in Module 4

Retrieval that pairs meaning-based similarity with keyword matching, because product codes, names, and exact phrases defeat embeddings alone.

Taught in Module 2

The property that makes a pipeline step safe to run again: once or five times, the result is the same. Without it, an automatic retry after a failure quietly duplicates data.

Taught in Module 3

Running a trained model to produce output. Training happens once; inference happens on every use, which is why it is where the running costs live.

Taught in Module 1

A management system standard certifying that an organisation runs AI deliberately: risks assessed and owned, decisions documented, performance reviewed.

Taught in Module 5

A crafted request designed to talk a model past its trained-in refusals. The reason output checking never relies on the model’s own judgement alone.

Taught in Module 4

A second model that reads an exchange and scores it against policy. The most capable check available, and the most expensive, so it is spent selectively.

Taught in Module 4

Lake storage underneath, with warehouse-style structure and guarantees applied only to the data that needs them. The default pattern for new analytical builds.

Taught in Module 3

A system trained on a vast amount of text to predict the next token. Its compressed patterns reproduce grammar, style, and knowledge, along with confident errors where the patterns run out.

Taught in Module 1

The ability to trace where a number came from and what transformed it on the way, so an auditor’s question has an answer.

Taught in Module 3

Software that learns patterns from data rather than following hand-written rules. Predictive machine learning scores and forecasts; generative AI is the branch that produces text and media.

Taught in Module 6

An open standard giving systems a common way to describe their tools to any model, replacing per-vendor integration the way one plug standard replaces a drawer of cables.

Taught in Module 4

The quiet decay of a predictive model as the world moves away from the history it learned. Found by monitoring, never by waiting for complaints.

Taught in Module 6

A model whose weights are published for anyone to download and run, either on infrastructure you administer or on compute you rent for the purpose, trading supplier convenience for control.

Taught in Module 6

Models that learn from historical data to score or forecast: arrears risk, fraud, demand, churn. Evaluated by outcomes across affected groups and monitored for drift, not judged by fluency.

Taught in Module 6

Everything placed in front of a model for one request: the standing instructions, the question, and any supporting material.

Taught in Module 1

A supplier feature that processes the unchanging start of a prompt once and stores the result, so repeat requests skip re-reading it. It makes long standing instructions nearly free after the first call.

Taught in Module 4

The craft of writing and testing the instructions, examples, and supporting material placed in front of a model, including the defensive share.

Taught in Module 1

Instructions hidden inside content a system reads, planted to hijack its behaviour. The indirect form arrives through retrieved documents, which is why retrieval systems treat content as data, never instructions.

Taught in Module 4

The pattern of retrieving relevant passages from your own documents and passing them to the model with instructions to answer from that material and cite it.

Taught in Module 2

A model that spends extra computation working through a problem before answering, trading speed and cost for performance on hard tasks. Priced accordingly.

Taught in Module 1

A careful second pass over a retrieval shortlist, promoting passages that answer the question over passages that merely resemble it.

Taught in Module 2

Machines rented for your exclusive use, in a region you name, running a model you chose. The middle arrangement between a shared API and hardware of your own: you pick the model and pay for the capacity whether or not you use it, and someone else keeps the machines running.

Taught in Module 6

Delivering a reply token by token as it is generated rather than waiting for the whole answer. In tension with output checking, which wants the complete draft first.

Taught in Module 4

The standing instructions placed before every conversation. In production systems a large share of it is defensive: what to refuse, how to treat retrieved content, when to hand over to a person.

Taught in Module 4

The setting controlling how much randomness enters a model’s choice of next token. Low is predictable and repetitive; high is varied and surprising.

Taught in Module 1

A short fragment of text, the unit models read, bill by, and measure context windows in. English averages roughly three quarters of a word per token.

Taught in Module 1

The convention that turns model text into action: the model emits a structured request naming a tool and its inputs, and the surrounding system, not the model, executes it.

Taught in Module 4

Information without a fixed shape: documents, emails, transcripts, images. What warehouses were built to exclude and AI systems consume.

Taught in Module 3

A database built to answer one question quickly: which stored items are nearest to this one? The home of embeddings in a retrieval system.

Taught in Module 2

The accumulated cost of leaving a supplier. Managed by keeping the portable things portable: your corpus, your prompts, your golden question set.

Taught in Module 6

The billions of internal numbers adjusted during training that hold everything a model learned. Frozen when training ends, until someone retrains.

Taught in Module 1