Glossary

Computer-use agents control applications the way people do: by looking at the screen and acting on it. This glossary entry defines the term, explains the perception-action loop, and sets out where the capability fits and where it does not.

Agent-to-agent protocols let one AI agent find, delegate to, and receive results from another, even across vendors. This explainer covers what A2A does, how it relates to Model Context Protocol, and the posture enterprises should take while the standard matures.

Indirect prompt injection is the attack that matters most for agents: instructions planted in a document, email, page, or record that the system reads and obeys. This glossary entry defines it, shows the attack shapes, and lists the defenses that hold.

Server-side rendering generates a page's HTML on the server for each request so browsers and crawlers receive complete content immediately. This explainer covers how SSR works, how it compares with client-side and static rendering, its costs, and how modern frameworks mix strategies per route.

Structured output makes language models return data your software can use: JSON that matches a schema every time. This explainer covers how it works, why it matters, how providers enforce it, what still fails, and how to design schemas well.

Test-time compute means spending more computation when a model answers, not only when it is trained, through longer reasoning, multiple samples, search, and verification. This explainer covers the techniques, the economics, and when the trade pays off.

A grounded AI answer is one whose claims are supported by the sources it was given. This explainer covers what groundedness is, how it differs from accuracy and citation, how it is measured with judges and humans, and how to improve it in RAG systems.

Hybrid search runs keyword and vector retrieval together and merges the results. This explainer covers why each method misses what the other catches, how results are fused and reranked, and when hybrid is worth the added complexity.

Jailbreaking is the practice of getting an AI model to ignore its safety rules and behavior constraints. This explainer covers how jailbreaks work, how they differ from prompt injection, what they risk for businesses, and how to defend in layers.

Latency decides whether an AI feature feels responsive or broken. This explainer covers the metrics that matter, where delay accumulates across retrieval, models, and tools, how to set budgets by use case, and the techniques that reduce it.

LLM-as-a-judge uses a language model to grade outputs against rubrics at scale. This explainer covers how it works, where it is reliable, its biases, how to calibrate it against human judgment, and how to use it in evaluation pipelines and production monitoring.

Prompt caching lets model providers reuse work for repeated prompt prefixes, cutting cost and latency. This explainer covers how it works, what qualifies, how to structure prompts for hits, what savings to expect, and where it does not help.

Users ask questions the way they think, not the way documents are written. Query rewriting reformulates, expands, decomposes, and contextualizes queries before retrieval. This explainer covers the techniques, when they help, and how to evaluate them.

ReAct is the pattern of interleaving reasoning with actions and observations that underlies most AI agents. This explainer covers how the loop works, why it outperforms plain prompting for tool use, where it fails, and the controls production agents need.

An orchestrator agent plans a task, delegates pieces to specialized agents, tracks progress, and integrates results. This explainer covers how orchestrators work, the patterns they use, their failure modes, and when a single agent is the better choice.

BM25 is the ranking function behind most keyword search, and it remains a core part of AI retrieval systems. This explainer covers how it scores documents, why exact-term matching still matters next to embeddings, how to tune it, and where it falls short.

Build-operate-transfer is an engagement model in which a partner builds a dedicated team, operates it until it is mature, and transfers it to the client. This explainer covers the three phases, how BOT compares with staff augmentation and captive centers, pricing, and risks.

Chunking decides what a retrieval system can find and what the model reads. This explainer covers why documents are split, the trade-offs in size and overlap, structure-aware and semantic strategies, metadata, and how to evaluate chunking choices.

Data lineage in AI is the recorded history of where data came from, how it was transformed, and where it was used across training, retrieval, and outputs. This explainer covers why lineage matters for trust, audits, and debugging, what to record, and how to implement it.

Data poisoning corrupts the data an AI system learns from or retrieves, changing its behavior without touching the model directly. This explainer covers attack types across training and retrieval, business impact, detection, and defenses.

Differential privacy is a mathematical guarantee that results do not reveal much about any single person in a dataset. This explainer covers how it works, the privacy budget, where enterprises apply it in analytics and AI, and its trade-offs.

Federated learning trains a shared model across many devices or organizations while raw data stays where it is. This explainer covers how the training loop works, what privacy it does and does not provide, where enterprises use it, and its costs.

Graph RAG retrieves from a knowledge graph of entities and relationships as well as from text passages, answering questions that span documents. This explainer covers how it works, when it beats standard RAG, how graphs are built, and what it costs.

AI red teaming is structured adversarial testing that tries to make an AI system fail, leak, or misbehave before attackers and users do. This explainer covers what it tests, how it is run, and how findings feed defenses and evaluation.