F3: AI Glossary
Quick-reference glossary of the 50 most important AI/ML terms. Each term is tagged with the FROOT layer it belongs to: Foundations, Reasoning, Orchestration, Operations, or Transformation.
:::tip How to Use This Glossary
Use Ctrl+F to search for a specific term. Layers map to FrootAI modules:
- F = Foundations (F1, F2) — model mechanics & architecture
- R = Reasoning — prompts, RAG, grounding, deterministic AI
- O = Orchestration — agents, tools, frameworks, memory
- Op = Operations — infrastructure, deployment, monitoring, observability
- T = Transformation — fine-tuning, evaluation, responsible AI, alignment :::
:::info Layer Legend Quick Reference
| Code | FROOT Layer | Focus Area |
|---|---|---|
| F | Foundations | Model mechanics, architecture, math, tokenization |
| R | Reasoning | Prompts, RAG, grounding, guardrails, search |
| O | Orchestration | Agents, tools, frameworks, memory, delegation |
| Op | Operations | Infrastructure, deployment, monitoring, cost |
| T | Transformation | Fine-tuning, evaluation, responsible AI, alignment |
| ::: |
Terms A–Z
| Term | Layer | Definition |
|---|---|---|
| Agent | O | An LLM-powered system that can reason, plan, and take actions using tools. Unlike simple chat, agents operate in loops — observe → think → act → observe. |
| Alignment | T | The process of making AI models behave according to human values and intentions. Techniques include RLHF, DPO, and constitutional AI. |
| Attention | F | The mechanism that lets transformers weigh the relevance of every token against every other token. Self-attention is the core of all modern LLMs. |
| Autoregressive | F | A generation strategy where each new token depends on all previously generated tokens. GPT-family models are autoregressive — they generate left-to-right, one token at a time. |
| BPE | F | Byte-Pair Encoding. The tokenization algorithm used by most LLMs. Iteratively merges the most frequent adjacent byte pairs into single tokens. See F1. |
| Chain-of-Thought (CoT) | R | A prompting technique that instructs the model to show its reasoning step-by-step before giving a final answer. Dramatically improves accuracy on math, logic, and multi-step problems. |
| Chunking | R | Splitting documents into smaller segments for RAG retrieval. Strategies include fixed-size (512 tokens), semantic (by topic), and recursive (by structure). Chunk size profoundly affects retrieval quality. |
| Context Window | F | The maximum number of tokens a model can process in a single request (input + output). GPT-4o: 128K, GPT-4.1: 1M. Exceeding it causes silent truncation. |
| Copilot | O | Microsoft's AI assistant brand. GitHub Copilot assists with code; Microsoft 365 Copilot assists with productivity. Built on GPT-4o with tool integration. |
| Cosine Similarity | F | A metric measuring the angle between two vectors (range: -1 to 1). Used to compare embeddings — 1.0 = identical meaning, 0 = unrelated. Core to semantic search and RAG. |
| Deterministic AI | R | Making AI outputs reproducible and predictable. Achieved via temperature=0, seed pinning, structured output schemas, and guardrails. See FrootAI Play 03. |
| DPO | T | Direct Preference Optimization. A simpler alternative to RLHF that fine-tunes models directly on human preference pairs without training a separate reward model. |
| Embeddings | F | Dense vector representations of text (e.g., 1536 or 3072 dimensions). Semantically similar texts have nearby vectors. Used for search, RAG, clustering, and classification. |
| Encoder/Decoder | F | Two transformer architectures. Encoders (BERT) create representations for understanding. Decoders (GPT) generate text autoregressively. Encoder-decoder (T5) does both. |
| Evaluation | T | Systematic measurement of AI quality. Key metrics: groundedness, relevance, coherence, fluency, safety. FrootAI uses automated eval pipelines with thresholds (≥4.0/5.0). |
| Few-Shot | R | Providing 2–5 input/output examples in the prompt so the model learns the pattern in-context. More reliable than zero-shot for formatting and classification tasks. |
| Fine-Tuning | T | Training a pre-trained model on domain-specific data to specialize its behavior. Cheaper than training from scratch. Methods: full fine-tuning, LoRA, QLoRA. See Play 13. |
| Foundation Model | F | A large pre-trained model (GPT-4o, Llama 3.1, Claude) designed to be adapted for many downstream tasks. Trained on broad data; specialized via prompting or fine-tuning. |
| Function Calling | O | An LLM capability where the model outputs structured JSON to invoke external functions/APIs. The model doesn't execute the function — your code does. Enables tool use. |
| GPU | Op | Graphics Processing Unit. The parallel-computing hardware that powers AI training and inference. Key metric: VRAM (memory). A100: 80GB, H100: 80GB, H200: 141GB. |
| Grounding | R | Connecting AI responses to verified source data (documents, databases, APIs) to reduce hallucination. RAG is the primary grounding technique. See Play 01. |
| Guardrails | R | Constraints applied to AI inputs and outputs — content filters, token limits, schema validation, blocklists. Implemented via Azure AI Content Safety or custom rules in guardrails.json. |
| Hallucination | R | When an AI generates plausible-sounding but factually incorrect information. Mitigated by grounding, RAG, low temperature, and groundedness evaluation. |
| Hybrid Search | R | Combining keyword search (BM25) with vector search (embeddings) for retrieval. Typically outperforms either alone. Azure AI Search supports this natively via search_type: "hybrid". |
| Inference | F | Running a trained model to generate predictions/outputs. What you do when you call an API. Contrast with training (learning weights from data). See F1. |
| In-Context Learning | R | The ability of LLMs to learn new tasks from examples provided in the prompt, without any weight updates. Encompasses zero-shot, few-shot, and many-shot prompting. |
| JSON Mode | R | A model setting that guarantees the output is valid JSON. OpenAI's response_format: { type: "json_object" }. More reliable: Structured Outputs with a JSON schema. |
| KV Cache | F | Key-Value cache. An optimization that stores previously computed attention keys and values to avoid recomputation during autoregressive generation. Reduces latency but consumes VRAM. |
| Knowledge Cutoff | F | The date after which a model has no training data. GPT-4o: Oct 2023. Information after this date requires RAG or tool use to access. |
| LangChain | O | A popular open-source framework for building LLM applications. Provides abstractions for chains, agents, tools, and memory. Python and JavaScript versions available. |
| LoRA | T | Low-Rank Adaptation. A parameter-efficient fine-tuning method that freezes the base model and trains small rank-decomposition matrices. Reduces VRAM by 10–100× vs full fine-tuning. |
| MCP | O | Model Context Protocol. An open standard (by Anthropic) for connecting AI models to external tools and data sources. FrootAI's MCP server exposes 25 tools. See F4. |
| Memory (Agent) | O | How agents persist information across turns. Short-term: conversation history in context. Long-term: vector store or database. Semantic Kernel uses ChatHistory + plugins. |
| Multi-Agent | O | Systems where multiple specialized AI agents collaborate on complex tasks. Patterns: supervisor, swarm, pipeline, debate. See Play 07 and Play 22. |
| Multi-Modal | F | Models that process multiple input types — text, images, audio, video. GPT-4o and Gemini are natively multimodal. Llama 3.2 Vision adds image understanding. |
| Next-Token Prediction | F | The core training objective of autoregressive LLMs. Given all preceding tokens, predict the probability distribution of the next token. This simple objective produces emergent capabilities. |
| ONNX | Op | Open Neural Network Exchange. A cross-platform model format for optimized inference. Used with ONNX Runtime for CPU/GPU deployment without framework dependencies. |
| Parameters | F | The learnable weights in a neural network. "7B" = 7 billion parameters. More parameters ≈ more capability but higher compute cost. See VRAM formula in F1. |
| Prompt Engineering | R | The practice of designing effective instructions for LLMs. Techniques: system prompts, few-shot examples, chain-of-thought, structured output, role-playing. |
| QLoRA | T | Quantized LoRA. Combines 4-bit quantization of the base model with LoRA adapters. Enables fine-tuning of 70B models on a single 48GB GPU. |
| Quantization | F | Reducing the numerical precision of model weights (FP32→FP16→INT8→INT4) to shrink VRAM usage and increase inference speed. Trade-off: some quality loss. |
| RAG | R | Retrieval-Augmented Generation. A pattern that retrieves relevant documents from a knowledge base and includes them in the LLM prompt for grounded answers. See Play 01. |
| RLHF | T | Reinforcement Learning from Human Feedback. A training technique where humans rank model outputs and a reward model is trained on those preferences to fine-tune the LLM. |
| Semantic Kernel | O | Microsoft's open-source SDK for AI orchestration. Supports plugins, planners, memory, and multi-model routing. The recommended orchestration layer for Azure AI apps. |
| Structured Output | R | Constraining LLM output to conform to a JSON schema. OpenAI's response_format: { type: "json_schema", json_schema: {...} } guarantees schema-valid output with 100% reliability. |
| Temperature | F | A generation parameter (0–2) controlling output randomness. 0 = greedy/deterministic, 0.7 = balanced, 1.5+ = highly creative. See F1. |
| Tokenization | F | The process of converting text into tokens (sub-word integers) that models can process. Different models use different tokenizers — tiktoken for OpenAI, SentencePiece for Llama. |
| Transformer | F | The neural network architecture (2017, "Attention Is All You Need") underlying all modern LLMs. Uses self-attention to process entire sequences in parallel. |
| Vector Database | R | A database optimized for storing and querying high-dimensional vectors (embeddings). Examples: Azure AI Search, Pinecone, Weaviate, Qdrant, pgvector. Core infrastructure for RAG. |
| Zero-Shot | R | Asking a model to perform a task with no examples — only instructions. Works well for capable models (GPT-4o) on common tasks. Falls back to few-shot when accuracy drops. |
:::info Didn't find your term?
The full FrootAI glossary in the MCP server covers 200+ terms. Run npx frootai-mcp@latest and use the lookup_term tool, or browse the Learning Hub.
:::
Common Confusions
| People Say | They Actually Mean | Correct Term |
|---|---|---|
| "The AI understands me" | Statistical pattern matching on tokens | Next-Token Prediction |
| "The model remembers" | Previous turns are re-sent in the context window | In-Context Learning |
| "Fine-tuning the prompt" | Iterating on the system/user message text | Prompt Engineering |
| "Open-source model" | Weights released but license may restrict commercial use | Open-Weight (usually) |
| "The AI is hallucinating" | The model generated ungrounded but plausible text | Hallucination |
| "RAG database" | A vector store used for retrieval-augmented generation | Vector Database |
Further Reading
- F1: GenAI Foundations — deep dive on tokens, parameters, and context windows
- F2: LLM Landscape — model families, pricing, and selection framework
- F4: GitHub Agentic OS — how agents, skills, and hooks work together