📘 Glossary of Key Concepts in GenAI and LM-Kit.NET

The LM-Kit.NET Glossary is a collection of the most essential terms and concepts related to Generative AI (GenAI) and Large Language Models (LLMs). It's designed to help developers, researchers, and enthusiasts navigate key ideas that shape text generation, model training, and inference in modern AI.

The goal here is to keep the explanations pedagogic, simple enough for beginners to grasp, yet detailed enough to be a good starting point for deeper exploration into GenAI. The glossary entries aim to break down complex topics in an easy-to-understand way, making it a useful resource for anyone looking to understand the core elements of LLM technology.

As LM-Kit.NET grows, expect regular updates with new terms and enhanced content to keep the glossary current and useful. It's a living resource that will evolve with the latest advancements in Generative AI and LLMs.

AI Agents & Orchestration

• AI Agents - Autonomous systems that integrate reasoning, logic, and tools
• AI Agent Delegation - Task assignment to specialized sub-agents
• AI Agent Execution - Runtime behavior and action processing
• AI Agent Grounding - Context, knowledge, and factual anchoring
• AI Agent Guardrails - Safety, validation, and control mechanisms
• AI Agent Memory - Persistent storage and recall across conversations
• AI Agent Orchestration - Multi-agent coordination and workflow patterns
• AI Agent Planning - Breaking down complex goals into actionable steps
• AI Agent Reasoning - Chain-of-Thought and logical inference
• AI Agent Reflection - Self-evaluation and iterative improvement
• AI Agent Skills - Modular, portable capability definitions via SKILL.md files
• AI Agent Tools - External functions that extend agent capabilities
• Function Calling - Dynamic method invocation during inference
• Model Context Protocol (MCP) - Standardized tool and resource access for agents

Model Architecture & Training

• Attention Mechanism - How models focus on relevant parts of input
• Context Windows - Maximum tokens a model can process at once
• Fine-Tuning - Adapting pre-trained models to specific tasks
• KV-Cache - Caching mechanism for efficient inference
• Large Language Model (LLM) - AI models trained to understand and generate text
• LoRA Adapters - Low-rank adaptation for efficient fine-tuning
• Mixture of Experts (MoE) - Sparse activation architecture with specialized expert networks
• Quantization - Model compression through reduced precision
• Small Language Model (SLM) - Compact, efficient language models
• Weights - Learned parameters in neural networks

Inference & Generation

• Chat Completion - Multi-turn conversation generation
• Dynamic Sampling - Neuro-symbolic adaptive inference framework
• Grammar Sampling - Constrained output generation
• Inference - Process of generating outputs from models
• Logits - Raw model output scores before normalization
• Perplexity - Measure of model prediction confidence
• Sampling - Token selection strategies (Top-K, Top-P, etc.)
• Speculative Decoding - Accelerated inference using draft model verification
• Symbolic AI - Rule-based reasoning and neuro-symbolic integration
• Text Completion - Single-turn text generation

Retrieval & Knowledge

• Embeddings - Vector representations of semantic meaning
• RAG (Retrieval-Augmented Generation) - Grounding responses in external data
• Reranking - Improving retrieval precision through scoring
• Semantic Similarity - Measuring meaning overlap between texts using vectors
• Vector Database - Storage and search for embedding vectors

Document Processing & Extraction

• Intelligent Document Processing (IDP) - End-to-end document understanding and automation
• Structured Data Extraction - Schema-based JSON extraction from content

Text Processing & Analysis

• Named Entity Recognition (NER) - Extracting entities from text
• Prompt Engineering - Designing effective inputs for language models
• Token - Basic units of text that models process
• Tokenization - Splitting text into tokens

Speech & Audio

• Voice Activity Detection (VAD) - Detecting speech in audio streams

Vision & Multimodal

• Vision Language Models (VLM) - Multimodal AI for image and text understanding