How to read this card. Terms are grouped by what you are doing, not alphabetical. If you are new: start with Core. If you build: skip to Prompting, Retrieval, and Agents. If you buy: Evaluation, Cost, and Risk are the ones that will show up in your contracts.
Core Concepts
| LLM | A model trained to predict text. What you are actually using when you say "AI" |
| Token | A chunk of text (about 3/4 of a word). What the model reads, and what you are billed by |
| Context window | The most text the model can consider at once. Its working memory |
| Embedding | Text turned into a list of numbers, so a computer can measure similarity |
| Inference | Running a trained model to get an answer. The thing you pay for |
| Parameters | The learned weights inside a model. Loosely: its size |
| Multimodal | Handles more than text: images, audio, video |
| Foundation model | A large, general-purpose base model that others build on |
Prompting
| Prompt | Everything you send to the model. Not just your question |
| System prompt | The standing instruction that sets role and rules for the whole session |
| Few-shot | Giving the model examples in the prompt to shape output |
| Chain-of-thought | Asking the model to reason step by step before answering |
| Structured output | A response constrained to a schema, usually JSON |
| Context engineering | Designing what goes into the window. Prompting is one part of it |
| Prompt caching | Reusing computed input to cut cost on repeated prefixes |
| Temperature | A dial for how much the model varies. Low for facts, high for ideas |
Retrieval (RAG)
| RAG | Retrieval-augmented generation: fetch relevant docs, then answer from them |
| Chunking | Splitting documents into passages small enough to embed |
| Vector database | Storage that finds text by embedding similarity, not keywords |
| Hybrid search | Combining keyword (BM25) and vector search. Beats either alone |
| Reranking | Re-scoring retrieved passages for relevance before the model sees them |
| Contextual retrieval | Adding LLM-generated context to each chunk before embedding. Cuts miss rate ~49% |
| Hallucination | A fluent, confident answer that is not supported by the source |
| Grounding | Forcing an answer to reference retrieved evidence |
Agents & MCP
| Agent | An LLM that plans, uses tools, and takes multiple steps toward a goal |
| Tool use | Letting the model call functions, APIs, or databases |
| MCP | Model Context Protocol. An open standard for connecting models to tools and data |
| Orchestrator | A lead model that decomposes a task and delegates to workers |
| Agent harness | The infrastructure around a model that lets it run long tasks reliably |
| Human-in-the-loop | A required human approval step inside an automated flow |
| ReAct | Reason, act (call a tool), observe, repeat. A common agent loop pattern |
| Function calling | A model's ability to output a structured request to invoke a specific function |