Glossary

Concise definitions of the terms used across the library. Each links to the document that explains it properly.

A–C

Agent — an LLM that takes actions through tools in a loop, observing results and deciding next steps, rather than answering in one shot. See Agents.

Attention — the mechanism by which each token gathers information from other tokens, weighted by learned relevance. See Attention Explained.

Base model — a pretrained model that only continues text, before instruction tuning. See What an LLM Actually Is.

BPE (byte-pair encoding) — a common subword tokenization scheme. See Tokenization.

Chain-of-thought (CoT) — prompting a model to reason in intermediate steps before answering. See Chain-of-Thought.

Chunking — splitting documents into passages for retrieval. See Chunking Strategies.

Context window — the bounded span of tokens a model can attend to at once; its working memory. See What an LLM Actually Is.

Cosine similarity — a measure of similarity between two vectors based on the angle between them. See Embeddings.

Cross-encoder — a model that scores a (query, document) pair jointly; used for reranking. See Retrieval and Reranking.

D–L

Decoding — turning the model's next-token probability distribution into concrete output tokens. See Sampling and Decoding.

Embedding — a vector representation of text where similar meaning means nearby points. See Embeddings.

Faithfulness (groundedness) — whether every claim in an answer is supported by its provided context. See Evaluating RAG.

Few-shot — including worked examples in the prompt to specify a task. See Few-Shot vs Zero-Shot.

Fine-tuning — further training a model's weights on task-specific data; good for changing behavior/style, not for injecting knowledge. See RAG End to End.

Hallucination — plausible but false model output; the next-token mechanism working as designed, not a separate bug. See What an LLM Actually Is.

Hybrid search — combining dense vector search with keyword (BM25) search. See Vector Databases.

In-context learning — a model inferring a task from examples in the prompt, without weight updates. See Few-Shot vs Zero-Shot.

Instruction tuning — training that makes a base model follow instructions as a helpful assistant. See What an LLM Actually Is.

Jailbreak — an attempt to bypass a model's safety behavior. See Jailbreaks and Defenses.

Knowledge cutoff — the point after which a model's training data ends and it knows nothing. See What an LLM Actually Is.

LLM-as-judge — using a model to grade another model's outputs. See LLM-as-Judge.

M–R

Mixture-of-Experts (MoE) — an architecture that activates only a few "expert" sub-networks per token. See The Transformer.

Prompt injection — untrusted text posing as instructions to hijack an LLM app. See Prompt Injection.

RAG (retrieval-augmented generation) — grounding a model in retrieved documents at query time. See Retrieval-Augmented Generation.

ReAct — an agent loop interleaving reasoning and tool actions. See ReAct and Tool Use.

Recall@k — the fraction of queries for which a correct item appears in the top k retrieved. See Evaluating RAG.

Reranking — reordering retrieved candidates by true relevance with a stronger model. See Retrieval and Reranking.

Residual connection — adding a sub-layer's input to its output; what makes deep transformers trainable. See The Transformer.

S–Z

Sampling — drawing the next token from the model's distribution (vs. always picking the top one). See Sampling and Decoding.

System prompt — the standing instructions and rules for an interaction, separate from per-turn user messages. See Prompt Patterns.

Temperature — a decoding knob controlling randomness/creativity. See Sampling and Decoding.

Token — the integer unit of text a model reads, usually a subword. See Tokenization.

Tool / function calling — a model emitting a structured, schema-conforming call for your code to execute. See ReAct and Tool Use.

Top-p (nucleus) sampling — sampling from the smallest set of tokens whose probabilities sum to p. See Sampling and Decoding.

Transformer — the stacked attention-plus-feed-forward architecture behind modern LLMs. See The Transformer.

Vector database — a store for fast (approximate) nearest-neighbor search over embeddings. See Vector Databases.

Zero-shot — specifying a task by instruction alone, with no examples. See Few-Shot vs Zero-Shot.

Return to the Reference section.

Grounding — tying a model's answer to retrieved source text so every claim can be traced to a citation.