AI Engineering for .NET glossary
Every term this track uses, defined once and linked from the lessons that spend it.
- Completion
- The text a model generates in response to a prompt, produced one token at a time. Also the name of the call that returns it, as opposed to an embedding call.See also Embedding, Token
- Context window
- The maximum number of tokens a model can attend to in one call. It covers the system prompt, the resent conversation, retrieved documents, tool schemas and the tokens about to be generated — input and output share it, so it is a budget rather than a memory.See also Token
- Deployment
- On Azure OpenAI, a named instance of a model inside your own subscription. Your code addresses the deployment name you chose rather than the model name, so the same configuration string means a different model in each environment unless you keep them aligned.See also Provider
- Embedding
- A fixed-length vector of floats representing a piece of text, where similar meanings land close together. An embedding call returns one and generates no text.See also Completion
- Grounding
- Supplying the facts an answer depends on in the prompt, rather than hoping the weights hold them. It is the standard containment for both fabrication and staleness, and it is what makes an answer auditable.See also Hallucination, Context window
- Hallucination
- A fluent, confident, specific and wrong answer. This course calls it fabrication, because nothing inside the model distinguishes it from a correct answer — both are high-probability continuations.See also Grounding
- Logit
- The raw score a model assigns to one vocabulary entry, before softmax turns the whole set of scores into probabilities. Every sampling option acts on logits or on the probabilities derived from them, which is why none of them changes what the model knows.See also Temperature, Token
- Prompt injection
- Instructions reaching the model through text you supplied as data — a retrieved document, a pasted stack trace, a user message. Every token in the context is treated alike, so this is a consequence of the architecture rather than a filtering problem, and the containments are architectural too.See also Grounding, Context window
- Provider
- The service that hosts a model behind an API — OpenAI, an Azure OpenAI resource, a vendor, or something running on your own machine. The same weights reached through two providers differ in auth, quota, residency and price, not in what they answer.See also Deployment
- Structured output
- Asking for a response that conforms to a schema and deserialising it, rather than parsing free text. The schema constrains the shape of an answer and never its truth, so domain validation is still yours to write.See also Completion, Hallucination
- Temperature
- A sampling parameter applied outside the model, controlling how far from the most probable token the sampler may stray. Zero makes the pick greedy; it does not make the response reproducible.See also Token
- Token
- A chunk of text from a fixed vocabulary learned by byte-pair encoding — roughly four characters of English prose, fewer of C#, far fewer of a GUID. It is the unit you are billed, timed and budgeted in.See also Context window
- Top-p (nucleus sampling)
- A sampling rule that keeps the smallest set of tokens whose probabilities sum to p, renormalises within that set and samples from it. It cuts the tail rather than a fixed number of candidates, so how many tokens survive depends on the distribution and not only on your setting.See also Temperature, Logit