Reference

AI glossary

Every term in the course, in one sentence each. Search it, or browse by topic to find out what you do not know yet. Free for everyone, forever.

You should not need this while reading. Any term in a lesson can be hovered or tapped for the same definition without leaving the page. This is here for when you want to look something up cold, or revise before a conversation you are nervous about.

How models work

ANN / HNSW / IVF: The indexing tricks that make searching millions of embeddings fast enough to be usable, by finding very close matches rather than guaranteed perfect ones.
Attention / self-attention: The mechanism letting every chunk of text weigh up which other chunks matter to it right now. This was the breakthrough that made modern models possible.
Chain-of-thought (CoT): Getting a model to work through a problem step by step before answering, which noticeably improves accuracy on anything with stages.
Context window: The maximum amount the model can hold in mind at once: your instructions, the conversation, any documents, and the answer being written.
Cosine similarity: Measuring how close two meanings are by the angle between their number lists. 1 means the same, 0 means unrelated.
Distillation: Training a small model to copy a big one, keeping much of the quality at a fraction of the cost and delay.
Embedding: A long list of numbers representing the meaning of a piece of text, arranged so that things meaning similar things end up close together.
Foundation / base model: A model that has only been pre-trained. Brilliant at continuing text, useless as an assistant, because nobody has taught it to answer questions yet.
Inference: Running a trained model to get an answer, as opposed to training it in the first place.
KV cache: Stored working-out that speeds up writing an answer. It gets larger as the input gets longer.
Logits / softmax: The raw scores the model gives every possible next chunk, and the step that turns those scores into percentages it can pick from.
Lost in the middle: Models read the beginning and end of a long input most carefully and can miss something buried in the middle.
Mixture of Experts (MoE): A design that sends each chunk of text to a few specialised sub-networks rather than the whole model, giving large capacity at lower running cost.
Multi-head attention: Running many attention operations side by side, each one drifting towards tracking a different kind of relationship, like grammar or what a pronoun refers to.
Open-weight vs closed model: Models you can download and run yourself, versus models you can only reach through somebody else's API. This decides everything when data cannot leave the building.
Parameters / weights: The numbers a model learned during training, quoted as 8 billion or 70 billion. More usually means more capable, and also slower and dearer.
Positional encoding / RoPE: How a model knows what order the words came in, given that it reads them all at once. Modern approaches are part of why long inputs became possible.
Pre-training: The expensive first stage where a model learns from a huge slice of the internet by predicting what comes next. This is where nearly all its knowledge comes from.
Prefill vs decode: The two phases of a request: reading your prompt, which sets how long you wait for the first word, and writing the answer, which sets how fast the rest arrives.
Quantization: Storing a model's numbers less precisely so it runs faster on cheaper hardware, at a small cost in quality.
Reasoning model: A model trained to think at length before answering. Much better at multi-step problems, and slower and dearer because you pay for the thinking.
Temperature: The randomness dial. Near zero gives consistent, factual, repetitive answers. Higher gives variety, creativity, and more risk.
Token: The chunk of text a model actually reads and writes, roughly three quarters of a word. You are billed per token and speed depends on them.
Tokenization / BPE: Chopping text into tokens using a fixed vocabulary. It was learned mostly from English, so other scripts often need two to four times more tokens for the same sentence.
Top-p / top-k: Two ways of ignoring the unlikeliest options before the model picks, so it never reaches for something bizarre.
Transformer: The design behind every modern language model. It reads the whole input at once rather than word by word.
Vector database: A store built to find the closest matching embeddings quickly, even across millions of documents.

Teaching a model

Abstention / fallback: Declining to answer when confidence or the retrieved evidence is too weak. Saying "I have no source for that" beats a confident invention every time.
Agentic RAG / GraphRAG: Letting the system search repeatedly and refine as it goes, or searching a map of how documents relate rather than just their text.
Catastrophic forgetting: When fine-tuning improves the task you trained on while quietly degrading everything else the model could do.
Chunking: Splitting documents into passages before storing them. Splitting along the document's own structure rather than by length is often the single biggest quality win.
Continued pre-training: Heavier domain adaptation, where you keep pre-training a model on a large body of specialist text before any fine-tuning.
DPO: A simpler, cheaper way of doing preference training that learns directly from pairs of better and worse answers.
Few-shot: Including a handful of worked examples in the request. The fastest way to pin down a format or style without retraining anything.
Fine-tuning: Continuing to train a model on your own examples so a behaviour becomes permanent. For how it behaves, almost never for what it knows.
Function calling / tool use: The model emits a structured request asking your code to run something. It never executes anything itself, which is where every safety control lives.
Grounding / citation: Making the model answer only from the sources you gave it, and point at which source each claim came from.
Hybrid search / BM25: Combining meaning-based search with old-fashioned keyword search. Meaning-based search alone will miss exact reference numbers and proper names.
PEFT / LoRA / QLoRA: Fine-tuning by training a tiny number of added parameters and freezing the rest. Roughly ninety-nine percent cheaper, and you can swap adapters in and out.
Prompting / in-context learning: Getting the behaviour you want purely through what you write in the request. Nothing is trained and nothing is remembered afterwards.
Query transformation / HyDE / multi-query: Rewriting or expanding the user's question before searching, so poorly worded questions still find the right documents.
RAG: Looking up the relevant passages in your own documents at the moment a question is asked, so the model answers from your material rather than from memory, with citations.
Re-ranking / cross-encoder: A sharper, slower model re-scoring the shortlist of retrieved passages so only the best few go to the model.
Reward model: A model trained to predict how a human would score an answer, used to guide the preference training stage.
RLAIF / Constitutional AI: Using a model, guided by written principles, to do some of the preference judging that humans would otherwise have to do.
RLHF: Teaching a model which answers people prefer, by having humans rank outputs and training the model towards the ones that score well.
RLVR / RFT: Training with rewards that can be checked automatically, like whether the maths is right or the code runs. Central to how reasoning models are built.
SFT / instruction tuning: Teaching a base model to follow instructions by showing it thousands of examples of an instruction and the ideal response.
Structured output / JSON mode: Forcing answers into a fixed shape so other software can read them reliably. This is what turns a chatty model into a component.
System prompt: The durable instructions, role, and rules, kept separate from whatever the user typed so your rules stay stable.

Agents

Agent: A model placed in a loop where it can take actions and see the results, repeating until a job is done. A chatbot answers, an agent acts.
MCP (Model Context Protocol): An open standard for connecting agents to tools and data. Think USB-C: one connector instead of a custom cable for every system.
Memory (short / long-term): Models forget everything between requests, so any memory is a system somebody built: the current conversation, or facts stored in a database and fetched back later.
Multi-agent orchestration / router / supervisor: Coordinating several specialised agents. A router sends work to the right one, a supervisor breaks a job up and delegates it.
ReAct: The most common agent pattern: alternate a thought about what to do next with an action that does it.

Measuring quality

Benchmark / saturation: Standardised exams for comparing models, and what happens when every model scores near the top so the exam stops telling you anything.
Component vs end-to-end eval: Testing the search step and the writing step separately, versus testing the whole chain. Component tests are what tell you where a bad answer went wrong.
Eval set / golden set: A curated set of realistic questions with known-correct answers, used to test the system. It is the exam, and a weak one gives confident numbers about nothing.
Faithfulness / groundedness: Whether every claim in the answer is actually supported by the sources retrieved. This is your detector for invented content.
LLM-as-judge: Using a strong model to grade outputs at scale, because humans cannot read ten thousand answers. It must be checked against human graders or it drifts.
Offline vs online eval: Testing against a fixed dataset before launch, versus measuring what actually happens with real users afterwards. You need both.
Position / verbosity / self-preference bias: The three known biases of a model grading other models: favouring whichever answer came first, favouring longer answers, and favouring its own family.
Precision / recall / F1: Of what the system flagged, how much was right. Of what it should have found, how much it found. And a single number balancing the two.
RAGAS metrics: The four standard measurements for a retrieval system: faithfulness, answer relevance, context precision, and context recall.
Regression testing: Re-running the whole test suite on every change, so a quality drop is caught before it ships rather than after somebody complains.

Data

Data flywheel: Real usage and human corrections feeding back into your tests and training, so the system improves faster the more it is used.
Human-in / on / out-of-the-loop: Three levels of oversight: a person approves before use, a person monitors and can intervene, or nobody is watching. Chosen by what a mistake costs.
Labelling / annotation / inter-annotator agreement: People attaching correct answers to raw data, and the measure of how often those people agree with each other. Low agreement means the task itself is not yet properly defined.
Red-teaming: Deliberately attacking your own system to find unsafe or wrong outputs before real users find them by accident.
RL environment: A simulated setting where a model practises a task repeatedly and is rewarded when it verifiably succeeds.
RLHF preference data: Human rankings of competing model outputs. The consistency of these shapes model behaviour more directly than almost anything else.
Synthetic data: Training or test material generated by a model. Useful for expanding coverage cheaply, and it reproduces the generating model's blind spots unless humans filter it.

Safety and security

Audit trail: A permanent log of what was asked, what sources were used, and what was answered. The difference between investigating a complaint and apologising for one.
Capability transfer / upskilling: Leaving an organisation's own people able to run and extend the system, rather than dependent on the supplier indefinitely.
Data residency / localisation: A legal requirement that data stays within national borders. It usually decides your entire architecture before anything else does.
Guardrails: The checks wrapped around a model, sitting in three places: on the way in, during generation, and on the way out.
Hallucination (intrinsic / extrinsic): A confident, fluent, false answer. Either contradicting the source you provided, or unsupported by any source at all.
Jailbreak: A prompt crafted to get around a model's safety training and produce something it should refuse.
PII / PHI: Personal and health information. Requires redaction, encryption, and above all sending the model no more of it than the task actually needs.
Prompt injection (direct / indirect): Malicious instructions hidden in text the model reads. The dangerous version hides them inside a document your system retrieves, so the attacker never touches your system directly.
RBAC / access control: Enforcing who can see what. It has to happen at the moment of retrieval, because filtering after the model has read something is not a control.
Sovereign AI: A country or organisation controlling its own data, models, and infrastructure, aligned to local law, language, and values.
VPC / single-tenant / on-prem / air-gapped: Where a system is allowed to run, from an isolated instance in the cloud through to a machine with no internet connection at all.

Running it in production

ASR / STT / TTS / WER: Speech to text, text to speech, and the standard measure of transcription accuracy. All three get harder outside English.
Drift: Real-world inputs or quality shifting over time, quietly, while your test scores stay the same.
Latency / TTFT / TPOT / p95: Response time, split into the wait before the first word and the speed of the words after it. Report the slow tail, not the average, because that is what people complain about.
LLMOps: The discipline of running AI in production: versioning, monitoring, cost, speed, safety, and continuous improvement.
Model routing / cascading: Sending easy work to a cheap fast model and escalating to an expensive one only when needed. Usually the biggest single cost saving available.
Observability / tracing: Recording the full path of every request so any single interaction can be investigated, debugged, or defended later.
Prompt caching: Reusing the computation for an unchanged opening section of a prompt, cutting both cost and the wait before the first word.
Quality-latency-cost triangle: The three-way trade-off in every deployment. You cannot maximise all three, so each piece of work gets placed deliberately.
Streaming: Showing each word as it is generated. It does not make anything faster, and it makes the wait feel as though it is already over.

Educational guidance

This is educational career guidance, not a promise of a job, a salary, or any specific outcome, and not financial or legal advice. Your results depend on your own effort and circumstances. Make big decisions with proper professional advice where it matters.