Resources
AI glossary
What these words mean, and — where it matters — where they get oversold. Written for someone reading a proposal and trying to work out what is actually being offered.
Terms
- Agent
A system built around a language model that can call tools — reading a database, updating a record, sending a message — rather than only producing text. What makes it an agent is the ability to act.
Where it gets oversold
The word is applied to everything from a scripted chatbot to a fully autonomous system. Ask what tools it can call and what it is allowed to change; the answer tells you which one you are being shown.
- Benchmark
A standard test set used to compare models. Useful for ranking models against each other in general.
Where it gets oversold
Benchmark performance rarely predicts performance on your data. A model topping a leaderboard can still be the wrong choice for your documents, which is why an evaluation set built from your own cases matters more.
- Chunking
Splitting documents into pieces small enough to retrieve and feed to a model. How you split them — by paragraph, by section, with or without overlap — substantially changes answer quality.
- Context window
The maximum amount of text a model can consider at once, including both what you send it and what it writes back. Measured in tokens.
Where it gets oversold
A large context window is not a substitute for retrieval. Filling it with everything you have is slower, more expensive, and often less accurate than sending the right two paragraphs.
- Drift
The gradual decline in a deployed model’s accuracy as the world changes away from the data it was trained on. Normal, expected, and the reason monitoring exists.
- Embedding
A numerical representation of a piece of text, such that similar meanings produce similar numbers. This is what makes it possible to search by meaning rather than by keyword.
- Evaluation seteval · evals
A collection of real cases with known-correct answers, used to measure whether a change made the system better or worse.
Where it gets oversold
If a supplier cannot tell you how they measure improvement, they are measuring it by impression. Ask to see the evaluation set.
- Fine-tuning
Further training an existing model on your own examples, so it produces output in a particular format, tone, or style.
Where it gets oversold
Frequently proposed as a fix for a model not knowing your facts. It usually is not — retrieval solves that better and cheaper. Fine-tuning is for shape, not knowledge.
- Hallucination
A model producing something fluent, confident, and false. Not a bug to be patched out — a property of how these systems work.
Where it gets oversold
Nobody can promise you a model that never does this. What can be promised is grounding answers in retrieved records so they are checkable, and logging them so a wrong answer can be traced.
- Idempotency
A property of an operation that produces the same result whether it runs once or five times. Essential in automation, where retries are normal.
Where it gets oversold
Without it, a retried step creates a duplicate order, a second invoice, or a repeated message. Any step that cannot safely run twice eventually will.
- Inference
Running a trained model to get an answer, as opposed to training it. Most of what an AI system costs to operate is inference.
- Large language modelLLM
A model trained on large amounts of text that predicts what text should come next. Everything they appear to do — answering, summarising, reasoning — is built on that.
Where it gets oversold
Understanding the mechanism explains the failure modes. A system that predicts plausible text will produce plausible text even when it has no basis for it.
- MLOps
The operational practice around deployed models: versioning, monitoring, retraining, and rollback. The part that decides whether a model is still accurate a year later.
- Model Context ProtocolMCP
An open standard for connecting models to tools and data sources through a consistent interface, rather than writing bespoke integration code per model.
- On-premiseon-prem · self-hosted
Running software, including models, on hardware you control rather than calling a third-party API. Usually chosen because data is not permitted to leave.
Where it gets oversold
It is a genuine option with a genuine cost: GPU hosting to pay for, and a lower quality ceiling than the largest hosted models. Anyone presenting it as strictly better is selling hardware.
- Open-weight modelopen source model
A model whose trained parameters are published, so it can be downloaded and run on your own infrastructure. Llama and Mistral are examples.
Where it gets oversold
"Open weights" is not the same as "open source" — the training data and process are usually not published, and licences vary in what commercial use they permit.
- Prompt
The instructions and context sent to a model to get a particular kind of output.
Where it gets oversold
Prompt engineering is real but limited. It cannot give a model facts it has no access to, and a system whose reliability rests entirely on prompt wording is fragile.
- RAGretrieval-augmented generation · retrieval
Finding relevant passages from your own documents and giving them to the model along with the question, so the answer is based on your material rather than on what the model absorbed during training.
Where it gets oversold
The retrieval step is where most of the quality comes from, and where most of the engineering effort should go. "We use RAG" says nothing about whether the right passage is being found.
- Structured output
Constraining a model to return data in a defined shape — JSON matching a schema — rather than prose that has to be parsed.
Where it gets oversold
If the next step is code, the model should be returning data. Parsing sentences with regular expressions is a reliable source of production bugs.
- Token
The unit models read and write in — roughly a short word or part of a word. Pricing and context limits are both measured in tokens.
- Tool usefunction calling
A model deciding to call a function you have provided — look up an order, check stock, send a message — and using the result in its answer. This is the mechanism underneath every useful agent.
- Vector database
Storage optimised for finding embeddings similar to a query embedding. What makes retrieval fast over a large document set.
Where it gets oversold
Often proposed earlier than needed. For modest document volumes, a PostgreSQL extension does the job without another system to operate.
Reading a proposal
Send it to us and we will tell you what it actually says.
No obligation, and no attempt to win the work off whoever wrote it. If it is a good proposal we will say so.