Skip to main content

Glossary of Key Terms

Link copied!

A jargon-buster for the terms used throughout this handbook. Where a full section exists, terms link to it.

Term Definition
AEO / GEO (Answer Engine / Generative Engine Optimization) Getting content surfaced and cited inside AI assistants' answers (ChatGPT, Perplexity, Google's AI answers) rather than ranked in classic search; the AI-era successor to SEO. The emerging llms.txt convention (a Markdown file pointing models at a site's key content) is one tactic; the term GEO comes from arXiv:2311.09735.
Agent An LLM given a goal, tools, and a loop. It plans, calls tools, observes results, and iterates until done. See Agentic Engineering.
Attention The transformer mechanism that lets each token weigh the relevance of every other token in context.
Chain-of-thought (CoT) Prompting or training a model to produce intermediate reasoning steps before its final answer, improving accuracy on hard problems.
Context window The maximum number of tokens a model can consider at once (prompt + output). Modern models range from ~128K to 1M+ tokens. See RAG.
Distillation Training a smaller "student" model to mimic a larger "teacher," which transfers capability at lower cost.
DPO (Direct Preference Optimization) A simpler, RL-free alternative to RLHF that tunes a model directly on preferred-vs-rejected response pairs.
Embedding A dense numeric vector representing the meaning of text (or image/audio); similar content sits close together in vector space.
Eval A structured test measuring model quality on a task or benchmark, used to compare models or guard against regressions.
Few-shot Including a handful of examples in the prompt to steer behavior. "Zero-shot" = no examples.
Fine-tuning Further training a base model on task- or domain-specific data to specialize its behavior.
Function calling / tool call The model emits a structured request (name + JSON args) that your code executes, returning the result to the model.
GGUF A file format for storing quantized model weights, used widely by llama.cpp and local-inference tooling.
Guardrails Input/output filters and policies that constrain a model by blocking unsafe content, enforcing formats, or checking for injection.
Hallucination Confident, fluent output that is factually wrong or fabricated. The core reliability problem of LLMs.
Harness The scaffolding around a model: the loop, prompt assembly, tool wiring, and control flow that turns a model into an application or agent.
Inference Running a trained model to generate output (as opposed to training it).
Jailbreak A prompt crafted to bypass a model's safety training and elicit restricted behavior.
KV cache Cached key/value attention tensors from prior tokens, which let generation reuse work instead of recomputing it. This is the main memory cost of long contexts.
Latency Time to produce a response. Often split into TTFT (time to first token) and per-token generation speed.
Logits The raw, unnormalized scores a model assigns to each possible next token before they're turned into probabilities.
LLM-as-judge Using a (often stronger) model to grade another model's outputs against criteria. Scalable but imperfect evaluation.
LoRA (Low-Rank Adaptation) A parameter-efficient fine-tuning method that trains small adapter matrices instead of full weights. QLoRA does this on a quantized base model for even lower memory.
MCP (Model Context Protocol) An open standard for connecting models to external tools, data, and resources through a uniform interface. See MCP.
MoE (Mixture of Experts) An architecture that routes each token to a few specialized sub-networks ("experts"). This gives large capacity at lower per-token compute.
Multimodal A model that handles more than one modality, e.g. text plus images, audio, or video.
Parameter A learned weight in a model. Count (e.g. "70B") roughly indicates capacity and cost.
PEFT Parameter-Efficient Fine-Tuning. Umbrella term for methods (LoRA, adapters, prefix-tuning) that adapt a model by training few weights.
Posttraining Everything done after pretraining to shape behavior: SFT, RLHF/RLAIF, DPO, etc.
Pretraining The initial, compute-heavy phase where a model learns from a massive unlabeled corpus by predicting the next token.
Prompt injection An attack where malicious instructions hidden in input (or fetched content) hijack the model's behavior.
Quantization Reducing weight precision (e.g. 16-bit → 4-bit) to shrink memory and speed inference, at some quality cost.
RAG (Retrieval-Augmented Generation) Retrieving relevant documents at query time and feeding them into the prompt so the model answers from grounded context. See RAG.
Reasoning / test-time compute Spending extra compute at inference (longer internal "thinking") to improve answers on hard tasks. This is the basis of reasoning models.
RLHF / RLAIF Reinforcement Learning from Human (or AI) Feedback. Aligns a model to preferences using a reward signal from human raters or another model.
SFT (Supervised Fine-Tuning) Fine-tuning on curated input→output demonstrations, typically the first posttraining step.
Speculative decoding Using a small fast "draft" model to propose tokens that the large model verifies in bulk. This accelerates generation.
Structured output Constraining a model to emit valid JSON (or a schema/grammar) so outputs are machine-parseable.
Temperature A sampling knob controlling randomness: low = focused/deterministic, high = diverse/creative.
Throughput Total work per unit time (e.g. tokens/second across all requests). A serving/capacity metric, distinct from latency.
Token The atomic unit a model reads and writes: a word-piece, roughly ¾ of a word in English. Pricing and limits are per-token.
Tokenizer The component that splits text into tokens (and back), defining the model's vocabulary.
Top-p (nucleus sampling) Sampling only from the smallest set of tokens whose cumulative probability exceeds p, which trims the unlikely tail.
TTFT (Time To First Token) Latency until the first output token appears; key for perceived responsiveness in streaming UIs.
Vector database A store optimized for similarity search over embeddings; the retrieval backbone of most RAG systems.
Zero-shot Asking a model to perform a task with no examples, relying purely on its pretrained/instruction-tuned knowledge.
Link copied!