Glossary
Glossary of AI and ML terms used across the wiki.
A quick reference for key terms used in AI Summary Hub.
| Term | Definition |
|---|---|
| A2A | Agent2Agent protocol — open standard (Linux Foundation) for agent-to-agent interoperability across frameworks and vendors. |
| Agentic RAG | RAG variant where the retrieval and generation steps are controlled by an autonomous agent that can iteratively plan and re-retrieve. |
| Agent | System that perceives and acts toward goals. |
| Computer Use | Capability allowing an LLM agent to control desktop GUIs (click, type, screenshot) as tools; Anthropic's term for its beta implementation in Claude. |
| Context Engineering | The practice of designing, managing, and optimizing what information is placed in an LLM's context window — including retrieval, compression, and caching strategies. |
| CoT | Chain-of-thought — step-by-step reasoning elicited via prompting. |
| DPO | Direct Preference Optimization — post-training technique that fine-tunes a model on human preference pairs without a separate reward model. |
| Extended / Test-time compute | Scaling inference compute (e.g. more search steps, longer thinking) rather than training compute to improve model output quality at test time. |
| GraphRAG | RAG variant that indexes documents as a knowledge graph, enabling multi-hop reasoning over entity relationships rather than flat vector retrieval. |
| GRPO | Group Relative Policy Optimization — RL-based post-training method (used in DeepSeek-R1) that improves reasoning by optimizing relative reward within a group of sampled outputs. |
| LLM | Large language model. |
| LLM-as-judge | Evaluation pattern where an LLM scores or ranks model outputs instead of (or alongside) human evaluators; used in automated evals and RLHF data pipelines. |
| MCP | Model Context Protocol — open standard (Anthropic) that defines how LLM hosts connect to external tools, data sources, and APIs via a client–server interface. |
| RAG | Retrieval-augmented generation — combining retrieval with LLM generation to ground responses in external knowledge. |
| ReAct | Reasoning + acting in interleaved steps — thought–action–observation loop for tool-using agents. |
| RDD | Retrieval-decision-design — spec-driven reasoning pattern that retrieves specifications, makes decisions aligned with them, and validates outputs. |
| RLHF | Reinforcement Learning from Human Feedback — post-training technique that uses human preference rankings to train a reward model, which then guides policy fine-tuning via RL. |
| Subagent | Agent within a hierarchical agent system, delegated sub-tasks by an orchestrator. |
| Transformer | Neural network architecture based on self-attention (e.g. BERT, GPT, Claude, Gemini). |
| Vibe coding | Informal term for AI-assisted programming where the developer describes intent in natural language and lets an LLM generate the implementation with minimal manual coding. |
See individual topic docs for detailed definitions and code examples.