Glossary of Key Terms

Every term introduced across the ten chapters of this tutorial, gathered in one place. Each entry links back to the chapter where it's explained in full, and to a real source where one exists.

Agent

An AI system wrapped in a loop that lets it take actions (calling a tool, reading the result and deciding what to do next) rather than only producing one reply. See Chapter 4.

Source: Yao et al., "ReAct" (2022)

API

A defined way for two programs to talk to each other over a network without either needing to know how the other is built internally. See Chapter 9.

Byte-Pair Encoding (BPE)

The subword tokenization algorithm behind every modern LLM's tokenizer: starting from individual characters, the most frequent adjacent pair anywhere in a training corpus is merged into a new symbol, repeatedly, until common words become single tokens and rare ones fall back to smaller pieces. See Chapter 1.

Source: Sennrich, Haddow & Birch, "Neural Machine Translation of Rare Words with Subword Units" (2015)

Cosine Similarity

A way of measuring how closely two embeddings point in the same direction, used to find the passages most relevant to a question. See Chapter 3.

Source: Cosine similarity, Wikipedia

Embedding

A list of numbers representing the meaning of a piece of text, positioned so that similar meanings end up close together. See Chapter 3.

Source: Mikolov et al., "Efficient Estimation of Word Representations in Vector Space" (2013)

Fine-Tuning

Further training an existing model on a smaller, specific dataset so it adapts to a narrower task, a different approach from RAG, which leaves the model unchanged and supplies context at answer time instead.

Source: Fine-tuning (deep learning), Wikipedia

Hallucination

A confident, fluent-sounding answer that is factually wrong, produced when a model is asked about something outside both its training and its current context. See Chapter 2.

Source: Ji et al., "Survey of Hallucination in Natural Language Generation" (2022)

HNSW

Hierarchical Navigable Small World: an indexing structure that lets a vector database skip most of its stored data and still find near-perfect matches quickly. See Chapter 9.

Source: Malkov & Yashunin, "Efficient and Robust Approximate Nearest Neighbor Search Using HNSW Graphs" (2016)

MCP (Model Context Protocol)

An open standard that lets any AI application discover and call tools exposed by any external service, without a custom integration for every pairing. See Chapter 4.

Source: Anthropic, "Introducing the Model Context Protocol" (2024)

Microservices

An architecture that splits an application into small, independent programs, each with one job, communicating over the network instead of sharing internal state. See Chapter 9.

ReAct

Short for "reason and act," the think, act, observe, repeat cycle an agent follows: decide what's needed, call a tool, read the result and decide again. See Chapter 4.

Source: Yao et al., "ReAct" (2022)

Retrieval-Augmented Generation (RAG)

A technique that searches a document collection for relevant passages and supplies them to a language model before it answers, instead of relying on the model's memory alone. See Chapter 3.

Source: Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020)

ROUGE-L

An evaluation metric that scores a candidate answer against a reference by the longest sequence of words they share, in order, a standard, cheap-to-compute signal used as one part of a larger evaluation pipeline. See Chapter 7.

Source: Lin, "ROUGE: A Package for Automatic Evaluation of Summaries" (2004)

Semantic Cache

A cache that reuses a stored result whenever a new request's embedding is close enough (by cosine similarity) to a previous one, instead of requiring an exact text match, catching paraphrased repeat questions a plain cache would miss. See Chapter 6.

Token

A small piece of text (often a whole word or a few characters) that is the basic unit a language model reads and generates, one at a time, and the unit every hosted model bills by. See Chapters 1 and 2.

Source: Vaswani et al., "Attention Is All You Need" (2017)

Tool

A defined capability (a calculator, a search function, a request to another system) that an agent is allowed to call, along with a description of what it does and what input it expects. See Chapter 4.

Source: Yao et al., "ReAct" (2022)

TextRank

A graph-based ranking algorithm that scores each sentence in a document by how similar it is to every other sentence, then keeps the highest-scoring ones as an extractive summary, the same random-walk idea behind PageRank, applied to sentences instead of web pages. See Chapter 5's extractive mode.

Source: Mihalcea & Tarau, "TextRank: Bringing Order into Text" (2004)

TF-IDF (Term Frequency)

A term-weighting scheme from information retrieval: words that occur often in a specific document, once common function words are filtered out, are a good signal of what that document is about. Chapter 5's extractive mode uses the "term frequency" half of this idea for its keyword extraction.

Source: Spärck Jones, "A Statistical Interpretation of Term Specificity and Its Application in Retrieval" (1972)

Vector Database

A database built to store embeddings and quickly find the ones most similar to a given query, typically using an approximate index like HNSW at scale. See Chapters 3, 8 and 9.

Source: pgvector, open-source vector search for Postgres