APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH R RAG: the short form of Retrieval-Augmented Generation (RAG). RAM: the short form of Random Access Memory (RAM). Random Access Memory (RAM): the fast working memory of a computer that holds what it is using right now. Rate limit: a cap on how many requests may be made within a set time. ReAct: a pattern that stands for Reasoning and Acting, in which an agent alternates between thinking and taking actions. Reactive pattern: a style in which the agent decides its very next step one step at a time with no upfront plan. Read-only tool: retrieves without changing anything. Recertification: the process of keeping a credential current. Red teaming: deliberately attacking your own system to find weaknesses before others do. Refactoring: reorganizing code to improve it without changing what it does. Reference implementation: a complete working example of one good way to build something. Regression: something that used to work breaking after a change. Regression test: a test that checks whether a new change broke something that used to work. Regression testing: checking that new changes do not break old behavior. Regulated field: an industry where rules control how work must be done. Release strategy: the plan for how a new version is delivered to users, such as slowly to a few users first. Replay: re-running a recorded session so it can be studied. Reporter: the harness component that produces the final summary of what happened in a run. Repository: a project folder whose changes are tracked over time. Repository root: the top-level folder of a repository. Reranker: a model that re-sorts search results by relevance. Resource and prompt: data to read and reusable templates offered by a server. Resource limit: a cap on things like memory, processing time, or disk space that a program may use. Responsible AI: building and using AI in ways that are fair, safe, honest, and accountable. Resume Review Agent: the small practice agent built in this book that reads fictional resumes and matches them to a fictional job posting. Retention period: how long records must be kept. Retrieval: finding and fetching the pieces of stored information that are relevant to a question. Retrieval-Augmented Generation (RAG): fetching relevant documents first and giving them to the model so its answer is grounded in real sources. Retry: trying a failed action again. Role: a label saying who is speaking in a message. Rollback: undoing steps to return to a known good earlier state. Root cause: the underlying reason a failure happened. Root-cause analysis: tracing a problem back to its true underlying cause instead of only fixing the symptom. Routing: sending each task to the right worker. Rubber-stamping: approving without real scrutiny. Rubric: a scoring guide that lists what to check and how many points each part is worth. Rug pull: a tool that changes behavior after being approved. Rule-based scoring: applying explicit coded checks. Rule-based system: software that follows rules written by people. Run: one complete attempt by an agent to do a job, from start to finish. Run identifier: a unique label attached to every record of a run. Runaway loop: a loop that keeps repeating without making progress and without stopping. Runbook: a guide for handling problems. Runtime: the time during which a program is actually running, or the environment it runs in. Runtime environment: the surroundings in which a program runs, with its resources and rules. Runtime execution environment: where actions are carried out.