Large Language Model
LLMAn extraordinarily well-read assistant that has no idea who you are until you tell it.
Imagine hiring a brilliant new analyst who has read essentially every book, manual, article, and webpage ever published. They are fluent in dozens of languages, can summarize anything, can write in any style you ask for, and can reason through complicated problems out loud. Now imagine they walk into your office on day one knowing absolutely nothing about your customers, your products, your processes, or yesterday's email. That is an LLM. ChatGPT, Claude, and Gemini are the brand names you have heard of. Underneath, they are all variations of the same idea: a model trained on a vast amount of text that has learned to predict what should come next in a sequence of words. The magic, and the limitation, is that they only know what you give them. The skill is not coaxing answers out of them. The skill is figuring out what to feed them and what to let them do with the answers.
Why it matters
When someone says "the AI got it wrong," nine times out of ten the real story is "we asked it something it had no way to know." Treat the LLM as a brilliant generalist that needs your information and your guardrails to be useful in your business.
See alsoPromptContext WindowRetrieval-Augmented Generation
Token
The little chunks of text that AI reads, writes, and charges you by.
When you read a sentence, you read words. When an LLM reads it, it sees something slightly different: pieces called tokens. A short common word like "cat" is usually one token. A long unusual word may be several. Punctuation and spacing can count too. Every question, document, answer, and reasoning step consumes tokens.
Why it matters
Tokens are the meter on the taxi. If a workflow reads ten long documents, reasons for a while, and writes a polished response, it costs more than a short classification task. A serious AI workflow should eventually know roughly what each run costs.
See alsoContext WindowPrompt
Context Window
How much the AI can hold in its working memory before something has to fall out.
The context window is the model's working desk. It can only reason over the material currently sitting on that desk. If you pile too much onto it, the important evidence can get buried or pushed out. Long documents, long chats, research packets, and customer records all need structure so the model sees the right facts at the moment it makes a decision.
Why it matters
When AI forgets instructions, misses obvious facts, or gives a generic answer after reading a lot, the problem is often not intelligence. It is that the system did not manage the working memory well.
See alsoTokenRetrieval-Augmented GenerationPrompt
Embedding
A way for software to notice that two ideas are related even when the words are different.
Think of an embedding as a meaning fingerprint. It lets a system understand that a customer email about a delayed shipment may belong near past notes about late freight, even if the exact phrase never appears. This is how AI search can find related policies, tickets, notes, and examples without relying only on exact keyword matches.
Why it matters
Embeddings are what make internal knowledge feel findable. They help a system retrieve the material a good employee would have remembered to check.
See alsoVector DatabaseRetrieval-Augmented Generation
Prompt
The instructions and context you give the AI. Often the cheapest lever to improve quality.
A prompt is not just the question typed into a chat box. In a production workflow, the prompt explains the job, the business role, the data boundaries, the format, the examples, the allowed uncertainty, and when the system should ask for human review. A good prompt is closer to a work instruction than a clever sentence.
Why it matters
Before spending money on new tools or custom training, inspect the prompt. Vague instructions produce vague systems. Clear operating instructions often produce a major quality jump.
See alsoLarge Language ModelEvalGuardrails
Fine-Tuning
Specialized training that changes how a base model behaves.
Fine-tuning teaches a model to produce a style, format, or behavior from examples. It can be useful, but many business problems people call fine-tuning problems are really context, retrieval, prompt, or workflow problems. If the AI needs current customer facts, fine-tuning is usually not the first answer.
Why it matters
Fine-tuning can be powerful, but it is also easy to overbuy. The right first question is: does the model need new behavior, or does it need better instructions and better business context?
See alsoRetrieval-Augmented GenerationPromptEval