Glossary - AI for Business

Key terms used throughout this book, listed alphabetically.

Agent — An LLM that can use tools (functions, APIs, databases) to take actions beyond text generation. Agents decide which tools to call based on the user’s question.

Attention — The mechanism that allows a transformer model to weigh the relevance of different parts of the input when producing each output token. Self-attention computes relationships between all pairs of tokens in a sequence.

Base Model — An LLM after pre-training but before any fine-tuning or instruction tuning. Base models are good at text completion but not at following instructions.

BPE (Byte Pair Encoding) — A subword tokenization algorithm that iteratively merges the most frequent pairs of characters or tokens. Used by GPT models.

Cascade (llm_cascade) — The automatic provider fallback system used in this book. When one LLM provider hits its quota, the cascade automatically tries the next available provider.

ChromaDB — An open-source vector database used for storing and searching embeddings. Used in several RAG chapters of this book.

CLIP (Contrastive Language-Image Pre-training) — A model by OpenAI that embeds both text and images into the same vector space, enabling cross-modal retrieval. Used in the Image RAG and Video RAG chapters.

Cosine Similarity — A measure of how similar two vectors are, computed as the cosine of the angle between them. Ranges from -1 (opposite) to 1 (identical). Used in RAG to find the most relevant document chunks for a query.

Cypher — A query language for graph databases, analogous to SQL for relational databases. Used with Kùzu in the Graph RAG chapter.

Dify — A visual LLM application builder that supports workflows, chatflows, knowledge bases, and tool integration. Used in the Deployment chapter.

Embedding — A dense vector representation of a token, word, sentence, or document. Embeddings capture semantic meaning: similar concepts have similar vectors.

FAISS (Facebook AI Similarity Search) — A library for efficient similarity search of dense vectors. Used in the LangChain chapter.

Fine-Tuning — Updating a pre-trained model’s weights on a specific dataset to teach it new knowledge or behavior. Contrasts with RAG, which retrieves knowledge at query time.

GGUF — A file format for quantized LLM weights optimized for CPU inference. Used by LM Studio and Ollama.

GloVe (Global Vectors) — A pre-trained word embedding model that captures semantic relationships from co-occurrence statistics. Used in the Embeddings notebook.

Google Cloud Run — A managed container platform on Google Cloud that autoscales from zero to many instances. Used in this book to host the MCP server as a public HTTPS endpoint. Each deploy ships a Docker container built automatically from the source folder.

Hallucination — When an LLM generates confident but factually incorrect information. RAG and tool use are designed to reduce hallucination by grounding answers in retrieved evidence.

Haversine Formula — A formula for computing the great-circle distance between two points on a sphere given their latitudes and longitudes. Used in the KU Parking Assistant.

JSON-RPC — A remote procedure call protocol encoded in JSON. The MCP protocol uses JSON-RPC 2.0 for communication between clients and tool servers.

LlamaHub — A registry of 300+ data loaders for LlamaIndex, supporting sources like Google Drive, Notion, GitHub, databases, and local files.

LLM (Large Language Model) — A neural network trained on vast text data to predict the next token. Examples: GPT-4, Gemini, Claude, Llama.

LM Studio — A desktop application for downloading, managing, and running open-weight LLMs locally with a visual interface.

LoRA (Low-Rank Adaptation) — A parameter-efficient fine-tuning technique that injects small trainable matrices into model layers, training only ~0.5% of parameters while keeping the rest frozen.

MCP (Model Context Protocol) — A standard protocol for exposing tools to LLMs. MCP servers provide tool discovery (tools/list) and execution (tools/call) over a standard interface.

Multimodal — Models or systems that process multiple types of input (text, images, audio, video) using the same architecture.

NF4 (NormalFloat4) — A 4-bit quantization format optimized for fine-tuning. Used in QLoRA to compress model weights to 25% of their original size.

Ollama — A command-line tool for running open-weight LLMs locally. Exposes an OpenAI-compatible API server.

Prompt Engineering — The practice of crafting input prompts to elicit desired behavior from an LLM. Includes techniques like few-shot examples, chain-of-thought, and system prompts.

QLoRA (Quantized Low-Rank Adaptation) — Combines 4-bit quantization with LoRA to enable fine-tuning of large models on consumer GPUs. Covered in the Fine-Tuning chapter.

Quantization — Compressing model weights from higher precision (FP16/FP32) to lower precision (INT4/INT8) to reduce memory usage and increase inference speed.

RAG (Retrieval-Augmented Generation) — A technique that retrieves relevant documents from a knowledge base and includes them in the LLM prompt, grounding the model’s answer in evidence rather than parametric memory.

ReAct (Reasoning + Acting) — An agent pattern where the LLM reasons about which tool to call, calls it, observes the result, and repeats until it has enough information to answer.

RLHF (Reinforcement Learning from Human Feedback) — A post-training technique where human evaluators rank model outputs, and a reward model trains the LLM to produce higher-ranked responses.

RunnableLambda — A LangChain wrapper that converts a plain Python function into a Runnable, enabling it to participate in LCEL pipe chains.

SFT (Supervised Fine-Tuning) — Training a base model on curated prompt-response pairs to teach it instruction following. The first step of post-training.

StateGraph — LangGraph’s core abstraction: a directed graph where nodes are processing functions and edges define the flow of a typed state dictionary.

System Prompt — Instructions given to an LLM that define its role, behavior, and constraints. Separate from the user’s message.

TF-IDF (Term Frequency-Inverse Document Frequency) — A classical text representation that assigns weights to words based on their frequency in a document relative to the corpus. High weight = important to that specific document.

Token — The fundamental unit of text that an LLM processes. A token may be a word, subword, or character, depending on the tokenizer.

Transformer — The neural network architecture behind all modern LLMs. Introduced in the 2017 paper “Attention Is All You Need.” Uses self-attention to process sequences in parallel.

TypedDict — A Python typing construct that defines a dictionary with specific key names and value types. Used in LangGraph to define the state that flows through a graph.

Vector Database — A database optimized for storing and searching high-dimensional vectors (embeddings). Examples: ChromaDB, Pinecone, FAISS, Weaviate, Milvus, LanceDB.

Vibe Coding — Building software by describing intent in natural language and letting an AI agent write the code. Coined by Andrej Karpathy in 2025.

Word2Vec — An algorithm that learns word embeddings by predicting context words. Demonstrates that semantic relationships are captured as vector arithmetic (e.g., king - man + woman ≈ queen).