01
Alignment
Definition
The process of making model behavior better match human goals, instructions, and safety expectations.
Back to Agentic AI Glossary
Agentic AI Glossary
LLM Core & Architecture Terms
LLM Core & Architecture Terms terms and explanations from the Agentic AI Glossary.
68 terms in this chapter
02
Assistant Message
Definition
A model-generated response stored as part of the conversation context.
03
Audio Model
Definition
Audio Model is a model type or component focused on audio. It helps produce, transform, rank, interpret, or evaluate AI outputs.
04
Chat Model
Definition
A model optimized for multi-turn conversational interaction.
05
Completion
Definition
The model-generated continuation, answer, or output produced from a prompt.
06
Completion Model
Definition
Completion Model is a model type or component focused on completion. It helps produce, transform, rank, interpret, or evaluate AI outputs.
07
Context Window
Definition
The amount of text, messages, tool outputs, and retrieved information a model can consider in one request.
08
Embedding Model
Definition
A model that converts data into vectors for search, clustering, memory, and similarity comparison.
09
Grounding
Definition
Connecting an AI answer to verified evidence, retrieved context, citations, or system-of-record data.
10
Hallucination
Definition
An incorrect or unsupported model output that sounds plausible but is not grounded in reliable evidence.
11
Instruction Following
Definition
A model's ability to obey user, system, and developer instructions accurately.
12
Logprobs
Definition
Probability scores for generated tokens, useful for confidence analysis and debugging.
13
Long-Context Model
Definition
Long-Context Model is a model type or component focused on long-context. It helps produce, transform, rank, interpret, or evaluate AI outputs.
14
Max Tokens
Definition
The maximum number of tokens a model is allowed to generate or process in a request.
15
Model Cascading
Definition
Using a sequence of models, often starting with cheaper models and escalating to stronger ones only when needed.
16
Model Distillation
Definition
Compressing knowledge from a larger model into a smaller model for faster or cheaper inference.
17
Model Routing
Definition
Selecting the best model for a request based on difficulty, cost, latency, risk, or required capability.
18
Model Selection
Definition
Choosing a model by matching task needs to accuracy, context length, modality, cost, latency, and safety requirements.
19
Multimodal Input
Definition
Input that includes more than one type of data, such as text plus images or audio.
20
Prompt
Definition
The input instructions, context, examples, or user request sent to a model.
21
Quantization
Definition
A compression technique that lowers model numeric precision to reduce memory and improve speed.
22
Reasoning Model
Definition
A model optimized to solve tasks that require multi-step analysis, planning, or problem solving.
23
Reranker Model
Definition
A model that reorders retrieved candidates by relevance to improve final context quality.
24
Speech-to-Text
Definition
A capability that converts spoken audio into written text.
25
Stop Sequence
Definition
A string or token pattern that tells generation when to stop.
26
System Prompt
Definition
High-priority instructions that define the model or agent role, boundaries, style, policies, and objectives.
27
Temperature
Definition
A sampling parameter that controls randomness and creativity in model outputs.
28
Text-to-Speech
Definition
A capability that converts written text into spoken audio.
29
Token
Definition
A unit of text processed by a language model, such as a word piece, character group, or symbol.
30
Tokenizer
Definition
The component that converts text into tokens and tokens back into text.
31
Top-k
Definition
A sampling method that limits token choices to the k most likely next tokens.
32
Top-p
Definition
A sampling method that limits token choices to the smallest set whose probabilities reach a threshold.
33
User Prompt
Definition
The user-provided message or request that initiates or guides the model response.
34
Vision-Language Model
Definition
Vision-Language Model is a model type or component focused on vision-language. It helps produce, transform, rank, interpret, or evaluate AI outputs.
35
Activation
Definition
The output value of a neuron or node after it processes its inputs.
36
Artificial Neural Network (ANN)
Definition
A computing system made of connected processing nodes inspired by biological neural networks.
37
Axon
Definition
A biological neuron structure that carries signals away from the neuron body to other neurons.
38
Back Propagation Algorithm
Definition
A supervised learning algorithm that updates neural-network weights by propagating error backward.
39
Bayes Net
Definition
Another name for a Bayesian network used to model probabilistic relationships.
40
Bayesian Network
Definition
A directed acyclic graph that represents probabilistic relationships among random variables.
41
Belief Network
Definition
Another name for a Bayesian network that reasons about uncertain domains.
42
Biological Neuron
Definition
A nerve cell that inspires the structure of artificial neural-network nodes.
43
Boolean Node
Definition
A Bayesian-network node that represents a proposition with true or false values.
44
Child Node
Definition
A node that is directly influenced by another node in a Bayesian network.
45
Conditional Probability
Definition
The probability of an event given that another related event or condition is known.
46
Conditional Probability Table (CPT)
Definition
A table that lists the probability of a node's values for each combination of parent-node values.
47
Content Addressable Memory
Definition
A memory style where information is retrieved by content similarity rather than by an explicit address.
48
Dendrite
Definition
A biological neuron branch that receives signals from other neurons.
49
Directed Acyclic Graph (DAG)
Definition
A directed graph with no cycles, often used to represent dependency structures.
50
Discrete Random Variable
Definition
A random variable that can take one value from a limited set of possible values.
51
Feedback ANN
Definition
A neural network topology where information can loop back through previous units.
52
FeedForward ANN
Definition
A neural network topology where information moves in one direction from input to output.
53
Instantiation
Definition
A specific assignment of values to variables or parent nodes in a probabilistic model.
54
Integral Value Node
Definition
A Bayesian-network node that represents integer-valued states such as age ranges or counts.
55
Link
Definition
A connection between neural-network nodes that carries a weighted signal.
56
Neurocomputer
Definition
A computing system designed around neural-network principles.
57
Node Value
Definition
The numeric output or activation value produced by a neural-network node.
58
Ordered Value Node
Definition
A Bayesian-network node whose possible values have a meaningful order, such as low, medium, and high.
59
Parent Node
Definition
A node that directly influences another node in a Bayesian network.
60
Pattern Classification
Definition
The task of assigning an input pattern to one of several categories.
61
Pattern Generation
Definition
The task of producing new examples that follow learned patterns.
62
Pattern Recognition
Definition
The task of detecting meaningful patterns in data such as images, speech, text, or signals.
63
Probabilistic Dependency
Definition
A relationship where one variable changes the likelihood of another variable.
64
Processing Element
Definition
A simple computing unit in a neural network that receives inputs, performs a calculation, and sends output.
65
Random Variable
Definition
A variable whose value is uncertain and described using probabilities.
66
Reinforcement Learning
Definition
A learning method where an agent improves decisions through rewards or penalties from interaction.
67
Supervised Learning
Definition
A learning method that trains from examples paired with known correct answers.
68
Unsupervised Learning
Definition
A learning method that finds hidden patterns in data without labeled answers.
Explore more chapters or test your knowledge with quizzes.