FAI Glossary
FAI Protocol terms, FROOT taxonomy, WAF pillars, and 200+ AI/ML terms — the complete reference.
FROOT Taxonomy
FROOT is the 5-layer knowledge architecture that organizes all AI concepts from fundamentals to production. Each letter maps to a learning layer, and each layer contains 2-6 modules that cover the full AI lifecycle:
Tokens, embeddings, model architectures, GPU compute. Modules F1 (GenAI Foundations), F2 (LLMs & SLMs), F3 (Glossary), F4 (.github Agentic OS). The bedrock layer — everything else builds on this.
Prompt engineering, RAG pipelines, deterministic AI patterns. Modules R1 (Prompt Engineering), R2 (RAG), R3 (Deterministic AI). The intelligence layer — how models think and retrieve knowledge.
Semantic Kernel, multi-agent systems, MCP tools, Azure AI services. Modules O1 (Semantic Kernel), O2 (AI Agents), O3 (MCP & Tools), O4 (Azure AI Foundry), O5 (Infrastructure), O6 (GitHub Copilot). The wiring layer — connecting models to the real world.
Azure AI Foundry, GPU infrastructure, Copilot integration. Shared with the Orchestration layer. The delivery layer — making AI solutions run at scale in production.
Fine-tuning, responsible AI, production readiness. Modules T1 (Fine-Tuning), T2 (Responsible AI), T3 (Production). The maturation layer — from prototype to enterprise-grade.
FAI Protocol Terms
The FAI terminology stack — from specification to distribution:
| Term | Definition |
|---|---|
| FAI Protocol | The context-wiring specification. Defined in fai-manifest.json. Connects agents, skills, instructions, hooks, plugins, and guardrails into a unified play. |
| FAI Layer | The conceptual glue between primitives. MCP handles tool calling, A2A handles delegation, AG-UI handles rendering — the FAI Layer handles wiring. |
| FAI Engine | The runtime that reads fai-manifest.json, resolves dependencies, wires primitives, runs evaluations, and reports status. |
| FAI Factory | The CI/CD pipeline — 20+ scripts and 15+ workflows that validate, generate, sync, and publish the ecosystem. |
| FAI Packages | Distribution channels: npm (frootai), PyPI (frootai-sdk), Docker, VS Code extension, GitHub Action. |
| FAI Toolkit | The 3 sub-kits: DevKit (agents, instructions, skills), TuneKit (openai.json, guardrails.json), SpecKit (schemas, validation). |
| FAI Marketplace | The discovery registry — 77+ plugins, searchable, installable via CLI or VS Code. Plugins are LEGO blocks that snap into any play. |
Play & Primitive Terminology
| Term | Definition |
|---|---|
| Solution Play | A complete AI solution template — manifest + primitives + infrastructure + config. Numbered 01-68+. |
| Primitive | A building block: agent, instruction, skill, hook, or plugin. Primitives are standalone but auto-wire when used inside plays. |
| Agent (.agent.md) | A scoped persona with YAML frontmatter — description, model, tools, WAF alignment. Invoked via @mention in Copilot Chat. |
| Instruction (.instructions.md) | Auto-attached context with applyTo glob patterns. Silently enriches Copilot for matching file types. |
| Skill (SKILL.md) | Multi-step workflow with tool access. Folder-based: skill-name/SKILL.md. Invoked via #skill reference. |
| Hook (hooks.json) | Event-driven automation. Fires on sessionStart, sessionEnd, userPromptSubmitted, preToolUse. Runs scripts automatically. |
| Plugin (plugin.json) | Packaged bundle of primitives. Installable from marketplace. Contains manifest, README, and all wired assets. |
Well-Architected Framework (WAF) Pillars
Every FAI solution play aligns to these 6 architectural pillars:
Managed Identity, Key Vault, RBAC, content safety filters, network isolation, secret scanning. Zero hardcoded credentials.
Retry with exponential backoff, circuit breakers, health checks, graceful degradation, multi-region failover.
Model routing (cheap model first, escalate if needed), token budgets, semantic caching, right-sizing compute.
CI/CD pipelines, IaC (Bicep/Terraform), structured logging, alerting, incident management, runbook automation.
Streaming responses, async processing, caching layers, batch operations, CDN for static assets, connection pooling.
Content safety, groundedness evaluation, fairness testing, transparency logging, human-in-the-loop, PII detection.
AI/ML Terms — Models & Architecture
| Term | Definition |
|---|---|
| LLM | Large Language Model — neural network with billions of parameters trained to predict text. GPT-4o, Claude, Llama. |
| SLM | Small Language Model — compact model (1-7B params) optimized for specific tasks. Phi-3, Mistral 7B. |
| Transformer | The attention-based architecture behind modern LLMs. Self-attention enables parallel processing of token sequences. |
| Token | The atomic unit of text for LLMs. ~4 characters in English. Both input and output are measured in tokens. |
| Context Window | Maximum tokens a model can process in a single request. GPT-4o: 128K tokens. Larger windows enable more context. |
| Embedding | Dense vector representation of text in high-dimensional space. Enables semantic similarity search. Key to RAG. |
| Fine-Tuning | Training an existing model on domain-specific data to improve task performance. LoRA, QLoRA, full fine-tuning. |
| LoRA | Low-Rank Adaptation — efficient fine-tuning that trains small adapter matrices instead of full model weights. |
AI/ML Terms — Inference & Generation
| Term | Definition |
|---|---|
| Temperature | Controls output randomness (0.0–2.0). Lower = deterministic, higher = creative. Use 0.0 for classification, 0.7 for chat. |
| Top-P | Nucleus sampling — selects from the smallest set of tokens whose cumulative probability exceeds P. Alternative to temperature. |
| Top-K | Limits sampling to the K most probable tokens. Reduces hallucination risk at the cost of diversity. |
| Hallucination | When a model generates plausible but factually incorrect content. Mitigated by RAG, grounding, and evaluation. |
| Grounding | Anchoring model outputs to verifiable source data. Groundedness score measures factual alignment. |
| Prompt Injection | Attack where malicious input tricks the model into ignoring instructions. Mitigated by content safety filters and input validation. |
| Streaming | Sending tokens to the client as they are generated instead of waiting for full completion. Improves perceived latency. |
| Semantic Caching | Caching LLM responses keyed by embedding similarity rather than exact string match. Reduces cost and latency for similar queries. |
AI/ML Terms — RAG & Retrieval
| Term | Definition |
|---|---|
| RAG | Retrieval-Augmented Generation — fetches relevant documents before generating, grounding responses in real data. |
| Chunking | Splitting documents into smaller segments for embedding indexing. Strategies: fixed-size, semantic, sliding window. |
| Vector Store | Database optimized for similarity search over embeddings. Azure AI Search, Cosmos DB vector, Pinecone, Weaviate. |
| Hybrid Search | Combining keyword (BM25) and vector similarity search for better recall. Azure AI Search supports this natively. |
| Reranking | Second-pass scoring of retrieved documents using a cross-encoder model. Improves precision after initial retrieval. |
| MCP | Model Context Protocol — open standard for connecting AI models to external tools and data sources. Replaces custom tool integrations. |
| A2A | Agent-to-Agent protocol — enables delegation between agents. Complements MCP (tools) and AG-UI (rendering). |
AI/ML Terms — Agents & Orchestration
| Term | Definition |
|---|---|
| Semantic Kernel | Microsoft's open-source AI orchestration SDK. Manages plugins, planners, and memory for agent applications. |
| Function Calling | LLM capability to select and invoke structured functions. The model outputs a JSON tool call instead of text. |
| ReAct | Reasoning + Acting pattern — the agent reasons about a step, takes an action, observes the result, then reasons again. |
| Chain-of-Thought | Prompting technique where the model shows its reasoning step by step. Improves accuracy on complex tasks. |
| Multi-Agent | Architecture where multiple specialized agents collaborate. Builder/reviewer/tuner is a 3-agent chain. |
| Guardrails | Runtime safety checks on model inputs and outputs. Content safety, PII detection, prompt injection resistance. |
| AG-UI | Agent-Generated UI protocol — standard for agents to render structured UI components (tables, charts, forms). |
Evaluation & Quality Metrics
| Metric | Scale | Definition |
|---|---|---|
| Groundedness | 1–5 | Are claims supported by the retrieved context? The most critical RAG metric. |
| Relevance | 1–5 | Does the response address the user's question? Measures answer alignment. |
| Coherence | 1–5 | Is the response well-structured and logically organized? |
| Fluency | 1–5 | Is the language natural and grammatically correct? |
| Safety | pass/fail | Does the response avoid harmful, biased, or inappropriate content? |