A-Z AI Glossary
Deep definitions, clear examples, and connected concepts for modern artificial intelligence terminology.
LLM
Large Language Model. Deep learning networks trained on vast textual datasets capable of generating and understanding natural text.
MCP
Model Context Protocol. An open-standard client-server protocol enabling LLMs to safely query data and access tools in external host environments.
RAG
Retrieval-Augmented Generation. Architecture that queries custom document databases to fetch factual context before querying the model.
Vector Embedding
A numerical vector representation of text or data capturing semantic conceptual relationships in high-dimensional space.
Fine-Tuning
Training a pre-trained model on custom datasets to alter its weights and customize its tone, formatting, or domain knowledge.
Inference
The stage where a trained model processes inputs (prompts) and generates outputs (completions).
Context Window
The maximum token payload a model can process in a single invocation, including prompt history and output.
AI Agent
An entity leveraging LLM reasoning loops, memory systems, and tools to execute tasks autonomously in an environment.
Tool Calling
The mechanism where an LLM formats a structured JSON command payload targeting an API, rather than generating natural language text.
Vector Database
A specialized database designed to store, index, and query vector embeddings using similarity algorithms (e.g. HNSW).
Multimodal
Models capable of handling multiple content formats (text, image, audio, video) natively as input and output.
Prompt
The instruction text or context passed to a language model to guide its completion output.
Temperature
A parameter scaling token probability distributions during inference, controlling output randomness/creativity.
Token
The basic unit of text processed by an LLM, typically representing a word, sub-word, or character sequence.
Tokenizer
A mathematical parser that converts raw text strings into arrays of token integer IDs matching a model’s vocabulary.
Hallucination
The phenomenon where an LLM generates factual untruths or nonsensical statements with high statistical confidence.
Quantization
Reducing the precision of model weights (e.g. converting float16 to int4) to reduce RAM size and increase inference speed.
LoRA
Low-Rank Adaptation. A parameter-efficient fine-tuning technique that injects small rank-decomposition matrices into transformer layers.
QLoRA
Quantized Low-Rank Adaptation. Running LoRA adapters on top of a base model quantized to 4-bit precision to save GPU memory.
Attention Mechanism
The mathematical layer in transformers calculating query-key relationships to weight token dependencies in text.
Transformer
The neural network architecture introduced in 2017 using attention mechanisms, replacing recurrent architectures (RNNs).
RLHF
Reinforcement Learning from Human Feedback. Aligning model behavior by training reward models based on human ranking preferences.
DPO
Direct Preference Optimization. An mathematical alternative to RLHF that optimizes policy networks directly using human preference datasets.
Cosine Similarity
A metric measuring the cosine of the angle between two multi-dimensional vectors, evaluating semantic alignment.
HNSW
Hierarchical Navigable Small World. A graph-based indexing algorithm used for fast approximate nearest neighbor searches.
IVF-Flat
Inverted File Flat. An indexing method partitioning vector space into clusters, reducing search scope to centroid zones.
Semantic Search
Retrieving relevant records by assessing conceptual intent and contextual meaning, ignoring exact keywords.
Dense Retrieval
RAG search using dense embeddings from neural encoders, capturing deep semantic relationships.
Sparse Retrieval
Keyword-based search relying on token frequency metrics (e.g. BM25), matching exact terms.
Hybrid Search
Combining dense vector similarity search with sparse keyword search to maximize retrieval accuracy.
Cross-Encoder
A deep-learning ranker scoring query-document pairs simultaneously to yield highly accurate relevance ratings.
Bi-Encoder
An embedder encoding queries and documents independently into vectors, allowing fast dot-product comparisons.
Re-ranking
The stage in RAG where candidate documents are re-ordered by relevance using high-precision ranker models.
Prompt Injection
A security exploit where user input overrides system prompt instructions to execute unauthorized commands.
Jailbreak
Constructing prompts that bypass a model’s safety alignment and guardrails to output restricted answers.
Guardrails
Verification layers filtering LLM inputs and outputs to ensure safety, formatting, and compliance.
Alignment
The training process of aligning AI behavior with human values, ethics, safety guidelines, and user intent.
Red Teaming
Systematically probing AI systems for security exploits, biases, and safety failures using adversarial prompts.
System Prompt
Core instructions defined by developers to establish an LLM’s role, rules, constraints, and operational context.
Zero-Shot
Prompting a model to solve a task without providing any input-output examples.
Few-Shot
Providing one or more input-output examples in a prompt to guide the model’s response formatting or reasoning.
Chain-of-Thought
A prompting technique instructing the model to break down its reasoning into step-by-step paragraphs before outputting answers.
ReAct
Reason + Act. An agent pattern combining step-by-step reasoning (Thought) with action triggers (Action/Observation).
Agentic Workflow
An iterative software loop where LLMs plan, fetch data using tools, evaluate outputs, and repeat until target states are met.
Short-Term Memory
Maintaining chat conversation history within the current context window to support dialogue coherence.
Long-Term Memory
Persisting user preferences, past interactions, or facts across separate sessions using vector stores.
Mixture of Experts
MoE. A transformer architecture routing inputs dynamically to specialized sub-networks (experts) during inference.
Parameter
The internal variables (weights and biases) of a neural network adjusted during the model training process.
Weights
The numerical values determining signal strength between connected nodes in neural network layers.
Backpropagation
The core training algorithm calculating loss gradients backwards through layers to adjust weights.
LLMOps
Large Language Model Operations. Best practices for managing the deployment, scaling, caching, and versioning of LLMs.
Data Drift
The decay of model accuracy over time as the distribution of production input data diverges from training data.
Latency
The time elapsed between sending an API request and receiving the first generated token (Time to First Token).
Throughput
The speed of generation measured in tokens per second processed by the serving server.
GPU
Graphics Processing Unit. Silicon processors optimized for massive parallel floating-point operations, powering modern ML training.
TPU
Tensor Processing Unit. Specialized ASIC chips designed by Google specifically to accelerate machine learning workloads.
NPU
Neural Processing Unit. Specialized silicon blocks inside consumer devices (like laptops or phones) designed for local inference.
GGUF
GPT-Generated Unified Format. A binary file format optimized for fast loading and running quantized models on consumer devices.
Llama.cpp
A lightweight open-source C/C++ runtime enabling high-speed LLM inference on diverse hardware (Mac, Windows, Linux).
CUDA
Compute Unified Device Architecture. NVIDIA’s proprietary software platform and programming interface for GPU computing.
ONNX
Open Neural Network Exchange. An open format built to represent machine learning models, allowing portability across runtimes.
Reranker
A specialized model that scores relevance of document-query pairs, used to filter candidates retrieved by vector search.
Instruction Tuning
Fine-tuning a base model on instruction-following datasets to convert it from a completion engine to an assistant.
Constitutional AI
An alignment framework (developed by Anthropic) training models to adhere to a predefined constitution of rules.
Self-Correction
An agent workflow where the model inspects its own code or output, identifies errors, and reprompts itself to fix them.
Model Drift
A decline in model accuracy or output formatting capability over time due to changes in base weights or system prompt changes.
Vector Store
A library or local module indexing vector arrays directly in memory or local files, avoiding a separate DB server.
Function Schema
A structured JSON schema defining function names, parameters, descriptions, and required arguments for tool calling.
Episodic Memory
Agent memory storing details of past execution steps and tool results during a specific task execution.
Semantic Cache
Caching API completions by checking if new prompts are semantically similar to past prompts, saving costs.
PagedAttention
An memory management algorithm for vLLM reducing VRAM waste in Key-Value caches by partitioning them like OS pages.
Model Server
A software runtime hosting models and exposing API endpoints (e.g. OpenAI compatible) to process requests.
Context Stuffing
Overloading a model’s prompt with excessive documents or text, which can degrade retrieval accuracy.
System Context
The set of background documents, databases, or variables loaded in an LLM application to support reasoning.
Top-P
Nucleus sampling. Filter tokens during inference based on the cumulative probability threshold P.
Top-K
Sampling from the top K highest probability tokens during inference, restricting output vocabulary scope.
Frequency Penalty
A parameter penalizing tokens based on how many times they have already appeared in the output, reducing repetition.
Presence Penalty
A parameter penalizing tokens if they have appeared in the output at least once, encouraging diverse topics.
Embeddings Model
A neural network designed specifically to generate compact vector representations of textual or media data.
TTS
Text-to-Speech. Models converting text strings into natural spoken audio waveforms.
STT
Speech-to-Text. Speech recognition systems converting audio speech inputs into text strings.
Diffusion Model
A generative model family creating images or media by iteratively denoising random Gaussian noise.
Latent Space
The compressed mathematical representation space where model features and concepts are organized.
Mixture of Agents
MoA. An agentic pattern where multiple model outputs are consolidated and refined by synthesizer agents.
Autoregressive
Generative models predicting the next token in a sequence based on all previously generated tokens.
Constitutional AI (Anthropic)
A method developed by Anthropic to train AI assistants to be helpful, harmless, and honest without human feedback.
Prompt Drift
A change in model response reliability due to updates in the model’s underlying weights or inference configurations.
Benchmarking
The practice of running standardized test batteries (e.g. MMLU, HumanEval) to measure model capability.
Context Recall
An evaluation metric assessing whether RAG retrieval systems successfully retrieved the correct document fact.
Context Precision
An evaluation metric assessing if the retrieved documents in RAG context are relevant to the query.
Faithfulness
An evaluation metric measuring whether the model’s answer is derived solely from the retrieved documents, avoiding hallucination.
Answer Relevance
An evaluation metric measuring whether the generated model response directly addresses the user query.
Model Evaluation
Assessing accuracy, latency, and security across model versions using programmatic test suites.
Model Registry
A centralized catalog storing, versioning, and managing weights and metadata of trained models.
Token Cost
The financial cost calculated based on prompt and completion token counts consumed during API requests.
Rate Limit
API request throttles restricting requests per minute (RPM) or tokens per minute (TPM) to prevent overload.
LPU
Language Processing Unit. A processor designed by Groq specifically to handle high-speed sequential token generation.
LoRA Adapter
A lightweight file containing custom adapter weight matrices trained via LoRA, loaded dynamically into a base model.
Embedding Space
The vector coordinate space where semantic embeddings are organized, allowing similarity computations.
Approximate Nearest Neighbor
ANN. Algorithms (e.g. HNSW, IVF) searching for vector matches within threshold distances quickly, sacrificing absolute accuracy.
Chunking
Splitting large documents into smaller semantic text blocks before vectorization, ensuring context chunks fit model inputs.
Semantic Similarity
Evaluating the closeness of two text segments based on conceptual meaning, computed via vector dot product or cosine.
Self-Attention
The core mechanism in transformers calculating relative token importances.
Multi-Head Attention
Running multiple attention heads in parallel to capture distinct features.
FlashAttention
An IO-aware exact attention algorithm optimizing GPU memory accesses.
Decoder-Only
Transformer model family predicting next tokens, optimal for generative chat.
Encoder-Only
Transformer model family producing dense embeddings, optimal for search.
Dense Vector
A numerical representation of text where every vector index contains non-zero float values.
Sparse Vector
A vector where most values are zero, optimal for term frequency scoring.
Inference Pipeline
The end-to-end execution flow of passing prompts and returning tokens in production.
Tokenizer Vocabulary
The set of unique token subwords recognized by a model tokenizer.
Greedy Decoding
Selecting the single highest-probability token at each generation step.
Model Grounding
The process of linking model outputs directly to verifiable external facts.
In-Context Retrieval
Retrieving documentation context to prepend directly inside prompts.
Hierarchical Indexing
Organizing vector spaces into parent-child clusters to optimize search speeds.
BM25
A lexical search relevance algorithm scoring document matches based on term frequency.
Reciprocal Rank Fusion
RRF. A ranking algorithm scoring combined retrieval lists based on position rankings.
Token Allocation
Budgeting prompt and completion contexts to prevent overflow.
Model Quantization
Compressing weights using lower bits to reduce runtime footprint.
Model Degradation
The drop in a model’s performance on standard benchmarks over time.
Adversarial Prompting
Constructing inputs designed to bypass model alignment safeguards.
Guardrail Layer
External validation software verifying prompts and completions before execution.
Structured Outputs
Forcing LLM outputs to adhere to strict JSON schemas.
Agent Memory Logs
Persistent history storage allowing agents to trace past executions.
Plan-And-Solve
An agent execution pattern dividing goals into milestones before running tools.
Human-in-the-loop
Designing manual gates requiring human approval for critical agent actions.
Rate-Limit Backoff
Retrying API calls exponentially after receiving 429 status codes.
Model Distillation
Training smaller student models on output targets of larger teacher models.
PEFT
Parameter-Efficient Fine-Tuning, adapting models with minimum parameter updates.
RLAIF
Reinforcement Learning from AI Feedback, utilizing model graders for preference scoring.
LoRA Rank
The dimension size of LoRA adapter weight matrices, regulating learning capacity.
RoPE
Rotary Position Embeddings, a position encoding technique enhancing long context windows.
Sliding Window Attention
Restricting attention scopes to local neighbor spans to accelerate decoding.
System Tracing
Logging trace spans of nested LLM pipelines for debugging.
Instruction Dataset
Curated prompt-response pairs used for training helper models.
LPU serving
Executing token inference on Groq specialized hardware arrays.
Headless Agent
An agent running asynchronously in background tasks without UI blockers.
Cosine Distance
Normalized vector distance metric calculating angles (1 – Cosine Similarity).