Glossary

45 terms in plain language, each linked to the module that teaches it.

Agent
A language model running in a loop that can decide to call tools and take multi-step actions toward a goal. Module 9
Agent2Agent protocol (A2A)
An open protocol that lets independent agents, possibly built on different frameworks, communicate as peers. Hosted by the Linux Foundation. Module 8
Barge-in
The ability of a voice assistant to be interrupted by the user while it is speaking. Module 19
BM25
A classic keyword-based ranking function used in search. Often combined with embeddings in hybrid retrieval. Module 6
Chat template
The exact format of role markers and special tokens an open-weight model expects. Use the tokenizer's template instead of building prompts by hand. Module 5
Chunking
Splitting source documents into passages that are indexed and retrieved. Good chunks follow the structure of the content. Module 6
Circuit breaker
A resilience pattern that stops calling a failing dependency for a while and uses a fallback immediately instead. Module 11
Code-switching
Mixing two or more languages within a conversation or sentence. Speech systems need vocabulary and detection that allow for it. Module 20
Containment rate
The share of conversations that end without escalation to a person. It can overstate success because it includes people who gave up. Module 17
Context window
The amount of text a model can consider at once, including instructions, conversation history, retrieved passages and tool results. Module 2
Embedding
A list of numbers representing the meaning of a piece of text, so that similar meanings are close together. Used for semantic search and routing. Module 2
Entity (slot)
A piece of information an assistant needs to collect or extract, such as a date or an order number. Module 2
Excessive agency
An OWASP LLM risk: an assistant has more functionality, permissions or autonomy than it needs and can take damaging actions. Module 18
Fallback ladder
An ordered set of recovery strategies, from retry and clarification to handoff, that changes approach as failures accumulate. Module 11
Function calling (tool use)
A pattern where a model returns a structured request to call a function, your code runs it, and the result goes back to the model. Module 8
Grounding
Tying a model's answer to provided sources or tool results so claims can be checked. Module 6
Guardrail
A check or constraint, in code or in the prompt, that limits what an assistant may say or do. Module 18
Hallucination
A fluent but unsupported or false statement produced by a language model. Module 6
Handoff
Transferring a conversation from an assistant to a person, ideally with a summary of context. Module 3
Idempotency key
A unique value sent with a write request so that retrying it cannot perform the action twice. Module 11
Intent
The goal behind a user's message in an intent-based system, such as changing a booking. Module 2
Journey ID
An identifier that follows a customer's task across channels so state, context and analytics stay connected. Module 21
LLM-as-judge
Using a language model with a rubric to grade another model's output. Needs calibration against human labels. Module 14
Locale
A language plus region and conventions, such as pt-BR or en-IN, written as a BCP 47 tag. It governs vocabulary, pronunciation and formats. Module 20
Metamorphic testing
Testing that a meaningful change to the input, such as a paraphrase or translation, does not change the expected behavior. Module 15
Model Context Protocol (MCP)
An open standard for exposing tools, resources and prompts to AI applications. Governed under the Agentic AI Foundation at the Linux Foundation. Module 8
Multi-agent system
An architecture with several specialized agents coordinated by a supervisor or through handoffs. Module 9
Natural language understanding (NLU)
Software that interprets language into structured meaning, typically intents and entities. Module 2
Open-weight model
A model whose trained weights are published, so you can host it yourself. Licenses vary, so read each one. Module 4
Prompt injection
An attack where attacker-controlled text changes a model's behavior, directly or through content the model reads. Module 18
Pronunciation lexicon (PLS)
A W3C standard file that defines how specific words are pronounced by a speech synthesizer. Module 20
Reciprocal rank fusion (RRF)
A simple method for merging several ranked lists by summing 1 divided by a constant plus each item's rank. Module 6
Reranking
Re-ordering retrieved candidates with a stronger model so the best passages reach the prompt. Module 6
Resolution rate
The share of conversations where the user's goal was actually met, measured by confirmation or a proxy such as no repeat contact. Module 17
Retrieval-augmented generation (RAG)
Searching your own content at question time and giving the best passages to a model to answer from, ideally with citations. Module 6
SSML
Speech Synthesis Markup Language, a W3C standard for controlling pronunciation, pauses and emphasis in synthesized speech. Module 20
State machine
A design with explicit states and transitions, where code decides what happens next and the model handles language. Module 9
Supersession
Marking a document version as replaced by a newer one, with dates, so retrieval uses only the current version while history is kept. Module 7
System prompt
The standing instructions that define an assistant's role, scope, rules and format. Treat it as a design document, and as extractable. Module 3
Test oracle
The mechanism that decides whether a test result is correct, such as an assertion, a schema, a rubric judge or a human reviewer. Module 15
Token
A piece of text, often part of a word, that a language model reads and produces. Costs and limits are usually counted in tokens. Module 2
Unbounded consumption
An OWASP LLM risk covering uncontrolled use of inference that causes denial of service or runaway cost. Module 18
Word error rate (WER)
The share of words a speech recognizer gets wrong. Measure it per locale and on names, numbers and intents, not only overall. Module 20