Glossary¶
A plain-language dictionary of every key term used across the course. Each entry gives a one-line definition, a slightly fuller explanation, and the chapter where it's introduced. Terms are grouped by theme; within each group they're ordered roughly from foundational to advanced.
Tip: when a chapter introduces a term, it links back here. This file is the single source of truth for definitions across the course.
Core search vocabulary¶
Information need — what the user actually wants to know, which they express imperfectly as a query. (Ch 1)
Query — the text a user submits to search. (Ch 1)
Document — a single retrievable unit of text (an article, paragraph, product description, etc.). (Ch 1)
Corpus — the full collection of documents you search over. (Ch 1)
Relevance — how well a document satisfies the user's information need. The thing search tries to maximize. (Ch 1)
Retrieval — the act of selecting candidate documents for a query from the corpus. (Ch 1)
Ranking — ordering retrieved documents so the most relevant appear first. (Ch 2)
Classical / lexical search¶
Lexical search — matching on the literal words (lexemes) in the query and documents. Also called keyword search. (Ch 2)
Tokenization — splitting text into smaller units (tokens), usually words or subwords. (Ch 2)
Stop words — extremely common words (the, a, of) often removed because they carry little meaning. (Ch 2)
Stemming / Lemmatization — reducing words to a base form ("running" → "run") so variants match. (Ch 2)
Inverted index — a lookup table mapping each term to the list of documents containing it; what makes keyword search fast. (Ch 2)
Term frequency (TF) — how often a term appears in a document. (Ch 2)
Inverse document frequency (IDF) — how rare a term is across the corpus; rare terms are more informative. (Ch 2)
TF-IDF — a classic score combining TF and IDF to weight terms by importance. (Ch 2)
BM25 — a refined, widely-used ranking function that improves on TF-IDF with length normalization and saturation. (Ch 2)
Vectors & embeddings¶
Vector — an ordered list of numbers; here, a point in a high-dimensional space. (Ch 3)
One-hot encoding — representing a word as a vector that is all zeros except a single 1. Sparse and meaning-blind. (Ch 3)
Bag-of-words — representing a document by its word counts, ignoring order. (Ch 3)
Dense vector — a compact vector where every dimension carries information (vs. sparse). (Ch 3)
Embedding — a learned dense vector that captures the meaning of text, so similar meanings sit close together. (Ch 3/4)
Embedding model — a neural network that maps text to an embedding. (Ch 4)
Dimensionality — the number of components in a vector (e.g., 384, 768, 1536). (Ch 4)
Cosine similarity — a measure of closeness based on the angle between two vectors; 1 = same direction, 0 = unrelated. (Ch 3)
Dot product — multiply-and-sum of two vectors; relates to similarity (equals cosine when vectors are normalized). (Ch 3)
Normalization — scaling a vector to unit length so only its direction matters. (Ch 3)
word2vec / GloVe — early models that learn one fixed vector per word. (Ch 4)
Contextual embeddings — vectors that depend on surrounding words (e.g., BERT), so "bank" differs by context. (Ch 4)
Semantic retrieval¶
Semantic search — finding documents by meaning rather than exact words. (Ch 5)
Dense retrieval — semantic search implemented by comparing dense embeddings. (Ch 5)
Bi-encoder — architecture that embeds query and documents separately, then compares; fast and scalable. (Ch 5)
Nearest neighbor search — finding the corpus vectors closest to the query vector. (Ch 5)
Approximate Nearest Neighbor (ANN) — algorithms that find almost the closest vectors far faster than exact search. (Ch 6)
Vector database — a system that stores embeddings and serves fast similarity search, with filtering and persistence. (Ch 6)
FAISS — a popular open-source library for efficient similarity search over vectors. (Ch 6)
HNSW — Hierarchical Navigable Small World; a graph-based ANN index, very common. (Ch 6)
IVF (inverted file index) — an ANN technique that partitions vectors into clusters to prune the search. (Ch 6)
Product quantization (PQ) — compressing vectors to save memory while keeping approximate distances. (Ch 6)
Improving results¶
Re-ranking — a second pass that re-scores top candidates more precisely. (Ch 7)
Cross-encoder — a model that reads query and document together for an accurate relevance score; slow but precise. (Ch 7)
Relevance score — the number a re-ranker assigns to a (query, document) pair: how well this document answers this query. Distinct from similarity, which only measures how alike two texts are. Valid for ranking within one query, not for comparing across queries. (Ch 7)
Hard negative — a training example that is wrong but tempting: topically close to the query and plausibly worded, yet not an answer (e.g. "the capital of Ontario is Toronto" for "what is the capital of Canada?"). Training on hard negatives — rather than random, easy ones — is what teaches a re-ranker the difference between similar and relevant. (Ch 7, 12)
Two-stage retrieval — retrieve many candidates cheaply, then re-rank a few expensively. (Ch 7)
Hybrid search — combining lexical (BM25) and semantic (dense) retrieval. (Ch 8)
Reciprocal Rank Fusion (RRF) — a simple, robust way to merge multiple ranked lists. (Ch 8)
Evaluation¶
Ground truth / relevance judgments — human-labeled "correct" results used to score a system. (Ch 9)
Precision@k — fraction of the top-k results that are relevant. (Ch 9)
Recall@k — fraction of all relevant documents found within the top-k. (Ch 9)
Mean Reciprocal Rank (MRR) — average of 1/(rank of first relevant result). (Ch 9)
nDCG — Normalized Discounted Cumulative Gain; rewards putting highly-relevant results near the top. (Ch 9)
MAP — Mean Average Precision across queries. (Ch 9)
Generation & RAG¶
Large Language Model (LLM) — a neural network trained to predict and generate text. (Ch 0/10)
Hallucination — when an LLM produces fluent but false or unsupported content. (Ch 10)
Retrieval-Augmented Generation (RAG) — retrieving relevant context and having the LLM answer using it, for grounded answers. (Ch 10)
Grounding — tying an answer to retrieved source text so it's verifiable. (Ch 10)
Context window — the amount of text an LLM can consider at once. Prompt and answer share it, so packing in more passages leaves less room to answer. (Ch 10)
Temperature — how adventurously an LLM samples its next token. Near 0 it almost always takes the most likely token (near-deterministic, repetitive); higher values give more varied and more creative — and more error-prone — output. Keep it low for factual RAG. (Ch 10)
Sampling / greedy decoding — generation draws each next token from a probability distribution (sampling), which is why the same prompt can give different answers on different runs. Always taking the single most likely token instead is greedy decoding, which is reproducible. (Ch 10)
Max tokens (truncation) — a hard cap on how many tokens a model may generate. Reaching it stops the output mid-sentence with no error — a common cause of "broken" RAG answers. (Ch 10)
Chunking — splitting documents into smaller pieces suitable for embedding and retrieval. (Ch 11)
Chunk overlap — repeating some text between adjacent chunks so context isn't cut off. (Ch 11)
Ingestion pipeline — the load → clean → chunk → embed → index process. (Ch 11)
Advanced¶
Query expansion / rewriting — improving a query by adding terms or rephrasing it. (Ch 12)
HyDE (Hypothetical Document Embeddings) — generate a hypothetical answer, embed that to retrieve. (Ch 12)
ColBERT / late interaction — multi-vector retrieval that compares token-level embeddings. (Ch 12)
Multimodal search — searching across modalities, e.g., text queries over images. (Ch 12)
Fine-tuning — further training an embedding model on your own labeled data. (Ch 12)
Agentic RAG — an LLM agent that plans multiple retrieval/reasoning steps. (Ch 12)