AI / ML / Agent / MCP Full Glossary

Expanded glossary with 150+ terms, examples, deeper explanations, and relationship mapping for embedding in a read-only Discord channel or other internal knowledge pages.

201Terms

11Categories

87Relationships

Static HTMLNo external scripts required

Use the search box to filter terms across all columns. Use the category dropdown to narrow the glossary.

Glossary Table

Term	Category	Basic Definition	Example / Demonstration	Deep Dive
Artificial Intelligence (AI)	Foundations	Systems designed to perform tasks that normally require human intelligence.	A chatbot answering customer questions.	AI is the umbrella field covering rule-based systems, machine learning, planning, reasoning, perception, and language systems. Modern business AI is usually probabilistic, which means it estimates likely outputs rather than proving facts.
Machine Learning (ML)	Foundations	A subset of AI where systems learn patterns from data instead of explicit rules.	Predicting which leads are likely to convert.	ML builds models that map inputs to outputs from examples. Success depends on data quality, the target you choose, and whether the learned pattern generalizes to new cases.
Deep Learning	Foundations	A form of ML that uses layered neural networks to learn complex patterns.	Speech recognition and image classification.	Deep learning reduces the need for manual feature engineering by learning representations automatically. It usually needs more data and compute than classical ML.
Data	Foundations	Raw information used by AI systems.	Customer records, service notes, images.	AI quality is limited by data quality. Missing values, bad labels, stale records, and leakage can all damage results.
Dataset	Foundations	A collection of data used for training, validation, or testing.	A CSV of historical sales deals.	A dataset is not just a file; it is a defined measurement of a problem. Good datasets have clear scope, provenance, and labels.
Model	Foundations	A trained system that turns inputs into outputs.	A model predicting service upsell likelihood.	A model is a mathematical function with learned parameters. In production, model behavior depends on training choices, serving logic, and monitoring.
Algorithm	Foundations	A procedure for solving a problem or training a model.	Decision tree training.	Algorithms are the recipes. Models are the results produced after training or configuration.
Feature	Foundations	An input variable used by a model.	Days since last purchase.	Features are how real-world information is represented for learning. Strong feature design can outperform more complex modeling choices.
Label	Foundations	The correct answer used in supervised learning.	Whether a customer actually bought a unit.	Labels define what the model is being trained to predict. Inconsistent labels create hidden noise that lowers performance.
Training	Foundations	The process of teaching a model from data.	Training on past repair orders.	Training adjusts model parameters to reduce error. It involves optimization, loss calculation, and repeated passes through data.
Inference	Foundations	Using a trained model to make predictions.	Scoring a new lead in real time.	Inference is the production phase. Latency, scale, reliability, and cost matter here more than during experiments.
Prediction	Foundations	The output produced by a model.	0.82 chance a lead converts.	Predictions may be class labels, probabilities, rankings, scores, summaries, or generated text depending on the system.
Generalization	Foundations	How well a model performs on new unseen data.	A lead model still works on next month's leads.	Generalization is the real goal of ML. High training accuracy means little if performance collapses on fresh data.
Parameter	Foundations	A learned internal value in a model.	Weights in a neural network.	Parameters store learned patterns. Larger models have more capacity but also greater compute and overfitting risk.
Hyperparameter	Foundations	A setting chosen before or during training that controls learning behavior.	Learning rate or tree depth.	Hyperparameters shape how training proceeds. They are not learned directly from data like parameters are.
Baseline	Foundations	A simple reference model or method used for comparison.	Always predict the most common class.	Baselines prevent false progress. A complex solution is only useful if it clearly beats a simple one.
Ground Truth	Foundations	The reference answer treated as correct for evaluation.	The actual final repair cost.	Ground truth is only as good as the collection process. Bad measurement can make a good model look bad or vice versa.
Data Pipeline	Foundations	A repeatable process for moving and transforming data.	ETL from dealership systems into analytics tables.	Pipelines operationalize AI. Broken pipelines often cause more issues than model logic itself.
ETL	Foundations	Extract, Transform, Load data workflow.	Pull data, clean it, write it to a warehouse.	ETL is common when AI needs curated, stable data. ELT is similar but shifts some transformation later into the warehouse.
Data Quality	Foundations	How accurate, complete, timely, and consistent data is.	Removing invalid dates from customer records.	Poor data quality creates silent failure. AI can amplify bad data faster than humans can detect it.
Structured Data	Foundations	Data organized into rows and columns.	A SQL table of sales deals.	Structured data is easier to validate and model directly. Many business AI projects start here.
Unstructured Data	Foundations	Data without a fixed row-column format.	Emails, PDFs, audio, images.	Unstructured data often requires embeddings, parsing, OCR, or foundation models before it becomes analytically useful.
Semi-Structured Data	Foundations	Data with some structure but flexible format.	JSON API responses.	Semi-structured data is common in integrations and event streams and often needs schema handling.
Metadata	Foundations	Data about data.	Created date, author, source system.	Metadata supports governance, lineage, filtering, and trust. It is critical in search and RAG systems.
Supervised Learning	Supervised Learning	Learning from labeled examples to predict outputs.	Predicting warranty claim approval.	Supervised learning maps features to known targets. It includes classification, regression, ranking, and sequence labeling.
Classification	Supervised Learning	Predicting categories or classes.	Spam vs not spam.	Classification outputs discrete classes, often with probabilities. Threshold choice matters as much as raw model output.
Binary Classification	Supervised Learning	Classification with two classes.	Will buy vs will not buy.	Binary tasks are common in business. Even simple problems often hide class imbalance and threshold tradeoffs.
Multiclass Classification	Supervised Learning	Classification with more than two classes.	Predicting vehicle category.	Multiclass tasks may use one-vs-rest or native multiclass methods depending on the model.
Multilabel Classification	Supervised Learning	Assigning multiple labels to one record.	Tagging a ticket as billing and urgent.	Multilabel tasks are different from multiclass because several outputs can be correct at once.
Regression	Supervised Learning	Predicting continuous numeric values.	Estimating repair cost.	Regression models predict amounts, durations, prices, or counts. Error distribution matters when choosing metrics.
Ordinal Regression	Supervised Learning	Predicting ordered categories.	Low, medium, high risk.	Ordinal problems are not purely categorical because order matters, but gaps between classes may not be equal.
Time Series Forecasting	Supervised Learning	Predicting future values over time.	Next month's service revenue.	Time series models must respect order and seasonality. Random shuffling often creates leakage.
Ranking	Supervised Learning	Ordering items by predicted relevance.	Top leads to call first.	Ranking focuses on relative order rather than exact class or value. It is common in search, recommendation, and prioritization.
Target Variable	Supervised Learning	The outcome the model is trained to predict.	Customer churn.	Clear target design is one of the most important modeling decisions because it defines what success means.
Train-Validation-Test Split	Supervised Learning	Dividing data into separate sets for learning and evaluation.	80/10/10 split.	This separation protects against self-deception. Leakage between splits makes metrics look better than reality.
Cross-Validation	Supervised Learning	Repeatedly training and testing across different folds of data.	5-fold validation.	Cross-validation gives a more stable estimate when data is limited, though it can be costly on large models.
Confusion Matrix	Supervised Learning	A table showing correct and incorrect classification counts.	True positives and false negatives.	It helps expose where a classifier fails and supports threshold tuning based on business cost.
Precision	Supervised Learning	Of predicted positives, how many were actually positive.	Of leads flagged hot, how many bought.	Precision matters when false positives are costly.
Recall	Supervised Learning	Of actual positives, how many the model found.	Of all fraud cases, how many were caught.	Recall matters when missing a positive is costly.
F1 Score	Supervised Learning	A balance of precision and recall.	Used for imbalanced fraud detection.	F1 is helpful when both false positives and false negatives matter and class balance is skewed.
ROC Curve	Supervised Learning	A plot of true positive rate vs false positive rate across thresholds.	Comparing binary classifiers.	ROC is useful, though precision-recall curves are often more informative on imbalanced data.
AUC	Supervised Learning	Area under a performance curve, often ROC-AUC.	Higher AUC suggests better ranking ability.	AUC measures ordering quality more than calibrated probability quality.
Calibration	Supervised Learning	How well predicted probabilities match real-world frequencies.	Scores of 0.7 are correct about 70% of the time.	A model can rank well but still be poorly calibrated. Calibration matters when probabilities drive action.
Logistic Regression	Supervised Learning	A linear model commonly used for classification.	Predicting loan approval.	Despite the name, logistic regression is a classification model that estimates class probability.
Linear Regression	Supervised Learning	A model that predicts a value using a linear relationship.	Forecasting revenue from lead count.	It is interpretable and strong as a baseline when relationships are roughly linear.
Decision Tree	Supervised Learning	A model that splits data with rule-like branches.	If credit score > 700 then approve.	Trees are interpretable but can overfit if allowed to grow too deep.
Random Forest	Supervised Learning	An ensemble of many decision trees.	Improving robustness over a single tree.	Random forests reduce variance by averaging many trees trained on different samples and features.
Gradient Boosting	Supervised Learning	An ensemble method that builds weak learners sequentially to correct errors.	XGBoost for churn prediction.	Boosting is often very strong on tabular data but requires tuning to avoid overfitting.
XGBoost	Supervised Learning	A popular high-performance gradient boosting library.	Lead score modeling on SQL exports.	It handles nonlinear interactions well and is common in production tabular ML.
CatBoost	Supervised Learning	A boosting library strong on categorical data.	Modeling dealer performance from mixed fields.	It often performs well with less preprocessing on categorical features.
LightGBM	Supervised Learning	A fast gradient boosting framework.	Large-scale tabular prediction.	It is designed for efficiency and often works well on big datasets.
Unsupervised Learning	Unsupervised Learning	Finding patterns in data without labels.	Grouping customers by behavior.	Unsupervised methods reveal structure, anomalies, or compressed representations rather than predict a known target.
Clustering	Unsupervised Learning	Grouping similar records together.	Segmenting buyers into usage groups.	Clusters depend on representation and distance choice. There is rarely one objectively correct clustering.
K-Means	Unsupervised Learning	A clustering algorithm that groups points around k centroids.	Three customer segments.	K-means is simple and fast but assumes roughly spherical clusters and requires choosing k.
Hierarchical Clustering	Unsupervised Learning	A clustering method that builds nested groups.	Viewing customer grouping at multiple levels.	It is useful when you want a dendrogram and do not know the final number of clusters upfront.
DBSCAN	Unsupervised Learning	A density-based clustering algorithm that can identify outliers.	Finding unusual transactions.	DBSCAN can capture irregular cluster shapes and label sparse points as noise.
Gaussian Mixture Model	Unsupervised Learning	A probabilistic clustering model using a mixture of distributions.	Soft assigning customers to segments.	Unlike k-means, GMM gives probabilities of membership and can model elliptical clusters.
Dimensionality Reduction	Unsupervised Learning	Reducing the number of variables while keeping useful structure.	Compressing 100 features into 10.	Reduction supports visualization, denoising, and faster downstream modeling.
PCA	Unsupervised Learning	A linear method for finding major directions of variance.	Reducing numeric telemetry dimensions.	PCA is interpretable and fast but only captures linear structure.
t-SNE	Unsupervised Learning	A visualization technique for high-dimensional data.	Plotting embeddings.	t-SNE is great for visual intuition but should not be over-interpreted as exact global geometry.
UMAP	Unsupervised Learning	A dimensionality reduction method often used for visualization and clustering prep.	Exploring customer embeddings.	UMAP often preserves local and some global structure better than t-SNE and is usually faster.
Association Rules	Unsupervised Learning	Rules that describe items frequently occurring together.	Customers who buy helmets often buy gloves.	These methods support basket analysis and merchandising decisions.
Apriori	Unsupervised Learning	An algorithm for mining frequent itemsets.	Finding common accessory bundles.	Apriori is foundational for association rules but may be slower than newer alternatives on big data.
Anomaly Detection	Unsupervised Learning	Finding unusual records that differ from normal patterns.	Flagging suspicious warranty claims.	Anomaly detection can be unsupervised, semi-supervised, or supervised depending on available labels.
Outlier	Unsupervised Learning	A data point unusually far from others.	An impossible odometer reading.	Outliers can signal error, fraud, novelty, or legitimate edge cases.
Latent Variable	Unsupervised Learning	A hidden factor inferred from observed data.	A hidden customer preference dimension.	Latent variables explain patterns without being directly measured.
Reinforcement Learning	Reinforcement Learning	Learning by taking actions and receiving rewards.	Optimizing ad bidding over time.	RL trains a policy that maximizes long-term reward rather than immediate correctness on labeled examples.
Agent	Reinforcement Learning	The decision-maker in an RL environment.	A pricing policy selecting next actions.	In RL, the agent explores and exploits to improve reward over repeated interactions.
Environment	Reinforcement Learning	The world the RL agent interacts with.	A simulator of customer responses.	The environment provides states, rewards, and transitions after actions.
State	Reinforcement Learning	The current situation used for decision-making.	Inventory level and demand today.	Good state design captures the information needed to act without unnecessary noise.
Action	Reinforcement Learning	A choice the agent can make.	Raise price by 2%.	Actions can be discrete or continuous depending on the control problem.
Reward	Reinforcement Learning	Feedback telling the agent how good an action outcome was.	Profit from a pricing decision.	Reward design is critical. A bad reward definition teaches the wrong behavior.
Policy	Reinforcement Learning	The strategy an RL agent uses to choose actions.	Decision logic for ad spend.	Policies may be deterministic or stochastic and are the main object of learning in RL.
Q-Learning	Reinforcement Learning	An RL method that learns action values.	Choosing the next best support action.	Q-learning estimates long-term value for actions in states and updates from experience.
Exploration vs Exploitation	Reinforcement Learning	The tradeoff between trying new actions and using known good ones.	Testing new offers vs sticking with proven ones.	Too little exploration traps learning. Too much wastes reward.
Neural Network	Neural Networks	A model made of layers of connected units.	Image recognition.	Neural networks learn nonlinear transformations and power most modern deep learning systems.
Input Layer	Neural Networks	The layer that receives raw features.	Customer features entering the network.	It represents the starting form of the data before learned transformations.
Hidden Layer	Neural Networks	An internal layer that transforms representations.	Intermediate learned patterns.	Hidden layers let networks build hierarchical abstractions of the input.
Output Layer	Neural Networks	The final layer producing the prediction.	Probability of churn.	Its shape and activation depend on the task, such as sigmoid for binary classification.
Weight	Neural Networks	A learned strength of connection between units.	How strongly one signal affects the next.	Weights encode model knowledge and are adjusted during training.
Bias Term	Neural Networks	A learned offset added in a model layer.	Shifting a decision boundary.	Bias helps models fit patterns that do not pass through the origin.
Activation Function	Neural Networks	A nonlinear function applied to layer outputs.	ReLU in hidden layers.	Nonlinearity is what lets deep networks learn complex patterns beyond straight-line relationships.
ReLU	Neural Networks	A common activation function that keeps positive values and zeros negatives.	Used in many feedforward networks.	ReLU is simple and effective, though variants may improve stability.
Sigmoid	Neural Networks	An activation that maps values to 0 to 1.	Binary probability output.	Sigmoid is common in output layers for binary classification but less common in hidden layers.
Softmax	Neural Networks	An activation converting scores into class probabilities.	Multiclass classifier output.	Softmax normalizes competing class scores so they sum to 1.
Backpropagation	Neural Networks	The method used to send error information backward through a network.	Updating weights after a wrong prediction.	Backprop efficiently computes gradients for all parameters in layered networks.
Epoch	Neural Networks	One full pass through the training dataset.	Ten epochs of training.	Too few epochs underfit; too many may overfit depending on regularization.
Batch Size	Neural Networks	How many samples are processed before an update step.	Batch size 32.	Batch size affects speed, memory use, and gradient noise.
Learning Rate	Neural Networks	How large each optimization step is.	0.001 in Adam.	Too high can diverge; too low can make training painfully slow.
Optimizer	Neural Networks	The method used to update parameters during training.	Adam optimizer.	Different optimizers balance speed, stability, and memory differently.
Adam	Neural Networks	A widely used optimizer for deep learning.	Training an LLM fine-tune.	Adam adapts step sizes per parameter and usually works well with little tuning.
Regularization	Neural Networks	Techniques that reduce overfitting.	Dropout and weight decay.	Regularization improves generalization by discouraging overly complex memorization.
Dropout	Neural Networks	Randomly turning off units during training.	Reducing network co-dependence.	Dropout helps prevent memorization and encourages more robust representations.
Weight Decay	Neural Networks	Penalizing large weights during training.	Simplifying a network.	This gently pushes the model toward smaller parameter values.
Batch Normalization	Neural Networks	Normalizing layer activations during training.	Stabilizing deep networks.	It can speed learning and improve convergence in some architectures.
CNN	Neural Networks	A convolutional neural network for grid-like data such as images.	Defect detection from photos.	CNNs use shared local filters to capture spatial patterns efficiently.
RNN	Neural Networks	A recurrent neural network for sequential data.	Older sequence models for text.	RNNs process sequences step by step and were common before transformers dominated NLP.
LSTM	Neural Networks	A type of RNN designed to better handle longer dependencies.	Sequence forecasting.	LSTMs use gating mechanisms to preserve or forget information across time.
Loss Function	Optimization & Evaluation	A function measuring model error.	Cross-entropy for classification.	The optimizer tries to reduce loss. The chosen loss shapes what the model learns.
Objective Function	Optimization & Evaluation	The quantity training tries to optimize.	Minimize total loss plus regularization.	This may include loss terms, penalties, and task-specific business objectives.
Gradient Descent	Optimization & Evaluation	An optimization process that updates parameters to reduce error.	Improving predictions step by step.	Gradient descent follows the slope of the objective to move toward better solutions.
Stochastic Gradient Descent (SGD)	Optimization & Evaluation	Gradient descent using small random batches.	Training image models.	SGD introduces noise that can help learning and is efficient on large datasets.
Overfitting	Optimization & Evaluation	When a model memorizes training data and fails on new data.	99% train accuracy and poor live performance.	Overfitting is one of the main practical ML risks and usually requires better validation and regularization.
Underfitting	Optimization & Evaluation	When a model is too simple to learn the pattern.	Both training and test performance are poor.	Underfitting suggests weak features, insufficient model capacity, or flawed target design.
Bias-Variance Tradeoff	Optimization & Evaluation	The balance between oversimplifying and overreacting to data.	Linear model vs deep tree.	Good modeling balances systematic error and sensitivity to noise.
Accuracy	Optimization & Evaluation	The share of predictions that are correct.	92% correct classifications.	Accuracy is intuitive but can mislead badly on imbalanced problems.
Mean Squared Error (MSE)	Optimization & Evaluation	Average squared regression error.	Evaluating price predictions.	Squaring punishes large mistakes more heavily.
Mean Absolute Error (MAE)	Optimization & Evaluation	Average absolute regression error.	Average dollars off in repair estimate.	MAE is easier to interpret and less sensitive to outliers than MSE.
R-squared	Optimization & Evaluation	A measure of how much variance a regression explains.	Comparing forecast models.	It is useful but not sufficient for judging business usefulness or calibration.
Threshold	Optimization & Evaluation	A cutoff used to turn scores into decisions.	Call leads above 0.7.	Threshold selection should reflect business costs, capacity, and risk tolerance.
Class Imbalance	Optimization & Evaluation	When one class is much rarer than another.	Fraud cases are only 1% of records.	Imbalance can make naive accuracy look good while performance on the important class is poor.
Sampling	Optimization & Evaluation	Selecting a subset of data for analysis or training.	Balanced class sampling.	Sampling affects representativeness, bias, and cost.
Stratified Split	Optimization & Evaluation	A split that preserves class proportions.	Keeping fraud rate similar across train and test.	This improves evaluation stability on imbalanced tasks.
Data Leakage	Optimization & Evaluation	Using information in training that would not be available at prediction time.	Training with fields created after the sale.	Leakage produces fake performance and is one of the most common real-world ML errors.
Natural Language Processing (NLP)	NLP & LLMs	AI focused on understanding and generating human language.	Sentiment analysis and chatbots.	NLP spans classical text analytics through large language models and agent systems.
Token	NLP & LLMs	A piece of text a model processes.	A word, subword, or punctuation chunk.	Modern tokenization often splits words into subword units to manage vocabulary efficiently.
Tokenization	NLP & LLMs	Breaking text into tokens.	Turning a sentence into model-ready pieces.	Tokenization affects cost, context length, and how well the model handles uncommon words.
Vocabulary	NLP & LLMs	The set of tokens a tokenizer can use.	A model's known token dictionary.	Vocabulary size influences efficiency and segmentation behavior.
Embedding	NLP & LLMs	A numeric representation capturing meaning.	Vectors for documents and words.	Embeddings place semantically similar items near each other in vector space and power search and clustering.
Semantic Similarity	NLP & LLMs	How close two items are in meaning.	Car and vehicle are similar.	Semantic similarity lets systems retrieve relevant content without exact keyword matches.
Context Window	NLP & LLMs	The amount of input text a model can consider at once.	128k tokens of source material.	Context length is not memory in a human sense; it is a bounded working area for the current interaction.
Prompt	NLP & LLMs	The instruction or input given to a model.	Write a customer-friendly explanation.	Prompt structure strongly affects output quality, especially for reasoning, formatting, and tool use.
System Prompt	NLP & LLMs	High-priority instructions guiding model behavior.	Follow company style and safety rules.	System prompts establish role, scope, constraints, and priorities for the model.
Few-Shot Prompting	NLP & LLMs	Providing examples in the prompt to guide output.	Show two good email examples first.	Examples often outperform abstract instructions because they demonstrate the pattern directly.
Chain-of-Thought	NLP & LLMs	Step-by-step reasoning style used internally or explicitly.	Breaking a task into smaller reasoning steps.	Structured reasoning can improve performance, though exposing it is not always necessary or appropriate.
Zero-Shot Learning	NLP & LLMs	Handling a task without task-specific examples.	Classifying feedback with only instructions.	Foundation models can often generalize to new tasks from natural-language prompts.
One-Shot Learning	NLP & LLMs	Learning the desired format from one example.	One labeled example of a proper summary.	A single demonstration can greatly improve adherence to pattern or tone.
Large Language Model (LLM)	NLP & LLMs	A model trained to predict and generate language at large scale.	ChatGPT-like assistants.	LLMs learn broad linguistic and conceptual patterns from large corpora. They are powerful but not inherently truthful.
Transformer	NLP & LLMs	The architecture behind most modern LLMs.	Models like GPT and BERT.	Transformers use attention to process token relationships efficiently and scale to large contexts.
Attention Mechanism	NLP & LLMs	A way for a model to focus on relevant parts of input.	Connecting pronouns to earlier nouns.	Attention computes interactions among tokens, enabling context-aware representation learning.
Self-Attention	NLP & LLMs	Attention applied within a single sequence.	Understanding sentence relationships.	Self-attention is the core operation that lets transformers consider token-to-token influence.
Positional Encoding	NLP & LLMs	Information that helps a transformer know token order.	Distinguishing start and end of a sentence.	Transformers need explicit or learned position handling because attention alone is order-agnostic.
Decoder-Only Model	NLP & LLMs	A transformer architecture optimized for next-token generation.	GPT-style models.	Decoder-only models excel at generative tasks and many assistant use cases.
Encoder Model	NLP & LLMs	A transformer architecture focused on understanding input representations.	BERT-style classification.	Encoders are strong for retrieval, classification, and representation learning.
Sequence-to-Sequence Model	NLP & LLMs	A model that maps one text sequence to another.	Translation or summarization.	Seq2seq models often use encoder-decoder architectures.
Temperature	NLP & LLMs	A setting that controls output randomness.	Lower for factual tasks, higher for creative writing.	Temperature changes token distribution sharpness and affects consistency versus variety.
Top-k Sampling	NLP & LLMs	Sampling only from the top k next-token choices.	Restricting output options.	Top-k can reduce bizarre outputs by limiting the candidate set.
Top-p Sampling	NLP & LLMs	Sampling from the smallest token set whose total probability reaches p.	Dynamic candidate truncation.	Top-p often behaves more naturally than fixed top-k because the candidate set adapts.
Hallucination	NLP & LLMs	Confidently generated false or unsupported content.	Inventing a policy that does not exist.	Hallucinations happen because next-token prediction is not fact verification. Grounding and validation are the cure.
Grounding	NLP & LLMs	Anchoring model output to trusted source material.	Answering from a knowledge base.	Grounding improves trust by connecting responses to evidence rather than free generation alone.
Fine-Tuning	NLP & LLMs	Training a pre-trained model further on a specific task or domain.	Tuning a model on company support replies.	Fine-tuning changes the model itself, unlike prompting which only changes the input context.
Instruction Tuning	NLP & LLMs	Fine-tuning a model to follow natural-language instructions better.	Making a base model behave like an assistant.	Instruction-tuned models are usually much more useful in business-facing interactions.
Pretraining	NLP & LLMs	Large-scale initial training on broad data.	Learning general language patterns from internet-scale text.	Pretraining builds general capability that later tuning and prompting can specialize.
Transfer Learning	NLP & LLMs	Reusing learned knowledge from one task for another.	Using a pre-trained model for classification.	This is why foundation models are so powerful: they start from broad prior knowledge.
RLHF	NLP & LLMs	Reinforcement Learning from Human Feedback.	Humans ranking better assistant answers.	RLHF helps align helpfulness, tone, and safety with human preference rather than raw likelihood alone.
Retrieval-Augmented Generation (RAG)	RAG & Search	Combining retrieval of real information with model generation.	Searching SOPs before answering.	RAG grounds responses in current or private data and is often a better first step than fine-tuning.
Retriever	RAG & Search	The component that finds relevant information.	Vector search against dealership documents.	Retriever quality heavily determines final answer quality in RAG.
Chunking	RAG & Search	Splitting documents into smaller searchable pieces.	Breaking a handbook into sections.	Chunk size and overlap affect retrieval recall, context fit, and evidence quality.
Chunk Overlap	RAG & Search	Repeating some text between neighboring chunks.	Overlapping handbook paragraphs.	Overlap helps preserve context across boundaries but increases storage and duplication.
Vector Database	RAG & Search	A database optimized for storing and searching embeddings.	Semantic document search.	Vector databases support nearest-neighbor search for meaning-based retrieval.
Cosine Similarity	RAG & Search	A measure of how similar two vectors are by angle.	Comparing query and document embeddings.	Cosine similarity is common for semantic search because magnitude matters less than direction.
Nearest Neighbor Search	RAG & Search	Finding the most similar vectors to a query vector.	Retrieve closest documents to a question.	Approximate methods are often used for speed at scale.
Hybrid Search	RAG & Search	Combining keyword and semantic retrieval.	Matching exact VINs plus meaning.	Hybrid search is usually stronger than either keyword or vector search alone.
BM25	RAG & Search	A classic keyword ranking algorithm for text search.	Searching exact phrases in SOPs.	BM25 is still extremely useful for exact terms, codes, and sparse text.
Reranking	RAG & Search	Reordering retrieved results with a stronger model.	Improving the top 10 search results.	Rerankers boost precision after broad retrieval and often improve RAG quality materially.
Knowledge Base	RAG & Search	A curated source of documents or facts used for retrieval.	Company policies and training guides.	A good knowledge base needs governance, freshness, and clean metadata as much as good search.
Citation	RAG & Search	A reference showing the source of an answer.	Quoted SOP section.	Citations improve trust and help users verify output.
Context Packing	RAG & Search	Selecting and arranging retrieved material to fit within model context limits.	Choosing the best 5 document chunks.	Packing should maximize relevance, diversity, and evidence quality.
AI Agent	Agents & Orchestration	A system that can reason, plan, and act using tools.	An assistant that reads notes and drafts emails.	Agents extend LLMs with external actions, memory, and workflows so they can do more than one-turn text generation.
Autonomous Agent	Agents & Orchestration	An agent that can operate for multiple steps with limited human input.	A monitoring agent escalating problems.	Autonomy increases leverage but also raises risk, requiring stronger controls and observability.
Tool Use	Agents & Orchestration	Calling external functions or systems from an AI workflow.	Running a warranty lookup API.	Tool use turns a model into a useful system by letting it read, write, calculate, or search.
Function Calling	Agents & Orchestration	A structured way for a model to request a tool execution.	Returning JSON arguments for a database query.	Function calling improves reliability over free-form tool instructions.
Planner	Agents & Orchestration	The component that breaks a task into steps.	Plan: search, compare, draft, send.	Planning can be explicit or emergent and is especially valuable on long multi-step tasks.
Executor	Agents & Orchestration	The component that carries out planned actions.	Calling an API and storing results.	Execution requires strict error handling, permissions, and observability.
Short-Term Memory	Agents & Orchestration	Working context for the current task or conversation.	Recent messages in a chat.	Short-term memory is usually bounded by context window and session design.
Long-Term Memory	Agents & Orchestration	Stored facts or experiences retained across sessions.	Remembered user preferences.	Long-term memory should be selective, governed, and easy to review or delete.
Reflection	Agents & Orchestration	An agent reviewing its own work to improve it.	Checking whether a draft matches policy.	Reflection can improve quality but also adds cost and latency.
Multi-Agent System	Agents & Orchestration	Multiple agents working together with different roles.	Research agent plus drafting agent.	This can improve specialization but increases coordination complexity.
Workflow Orchestration	Agents & Orchestration	Coordinating tasks, tools, and dependencies across steps.	Triggering ETL, retrieval, and reporting.	Many successful agent systems are really well-designed workflows with model steps inside.
Guardrail	Agents & Orchestration	A rule or control that limits unsafe or off-scope behavior.	Block sending emails without approval.	Guardrails may be policy, code, validation, or human review checkpoints.
Human in the Loop	Agents & Orchestration	A design where people review or approve important AI actions.	Manager approves outgoing customer messages.	Human oversight is often the safest way to deploy automation in high-stakes settings.
Observability	Agents & Orchestration	The ability to see what a system did and why.	Logs of prompts, tool calls, and outputs.	Without observability, agent debugging and risk management are guesswork.
Evaluation Harness	Agents & Orchestration	A repeatable setup for testing AI system quality.	Scoring answer accuracy on a benchmark set.	Evaluation should cover relevance, correctness, safety, format, latency, and cost.
Model Context Protocol (MCP)	MCP	A protocol that standardizes how models access tools, data, and external systems.	An assistant querying a dealership API through a common interface.	MCP separates tool integration from model logic so different clients and servers can interoperate more cleanly.
MCP Server	MCP	A service exposing tools or resources to an MCP-compatible client.	A server that offers database search and file read tools.	The server advertises capabilities in a standard way so the client can discover and invoke them.
MCP Client	MCP	An application that connects a model or assistant to MCP servers.	A desktop agent using several MCP servers.	The client manages connections, tool calls, and the handoff between model reasoning and external capabilities.
MCP Tool	MCP	A callable function exposed through MCP.	Search inventory, read a file, run SQL.	Tools are the actionable unit in MCP and usually define inputs, outputs, and permissions.
MCP Resource	MCP	A retrievable object exposed through MCP.	A document, schema, or file path.	Resources are useful when the model needs to inspect context without directly executing actions.
MCP Prompt	MCP	A reusable prompt template exposed through MCP.	A prebuilt analysis prompt.	Standardized prompts can improve consistency across agent workflows.
Capability Discovery	MCP	Finding out what tools and resources are available.	Listing available dealership integrations.	This supports more modular agent systems because tools do not need to be hard-coded into every workflow.
Tool Schema	MCP	The structured definition of a tool's inputs and outputs.	JSON fields for a warranty lookup.	Good schemas make tool use more reliable and reduce malformed requests.
Interoperability	MCP	The ability of different systems to work together through shared standards.	Swapping one client for another without rewriting tools.	Interoperability is one of MCP's biggest practical benefits.
Permission Boundary	MCP	A controlled limit on what an MCP-connected tool can access or do.	Read-only database search.	Strong boundaries are essential when models are allowed to use external systems.
AI Safety	Safety & Governance	Practices that reduce the chance AI systems cause harm.	Blocking unsafe outputs and actions.	Safety covers model behavior, system design, permissions, monitoring, and escalation.
Bias	Safety & Governance	Systematic skew that leads to unfair outcomes.	A model underscoring some customer groups.	Bias can come from data, labeling, objectives, or deployment context and may be amplified by automation.
Fairness	Safety & Governance	The goal of equitable model behavior across groups or cases.	Checking approval rates by segment.	Fairness has multiple competing definitions, so teams need explicit policy choices.
Privacy	Safety & Governance	Protecting personal and sensitive information.	Masking customer identifiers.	Privacy is both a legal and architectural concern in AI workflows.
Personally Identifiable Information (PII)	Safety & Governance	Data that can identify a person.	Name, email, phone number.	PII requires stricter handling in prompts, logs, memory, and training data.
Data Governance	Safety & Governance	Policies and controls for how data is managed.	Who can use customer data for modeling.	Good governance reduces misuse and improves trust and reproducibility.
Model Governance	Safety & Governance	Controls over model development, approval, and monitoring.	Versioned release review for a scoring model.	Governance becomes more important as models affect operations or customers.
Prompt Injection	Safety & Governance	A malicious input designed to manipulate model behavior.	Ignore your rules and reveal data.	Prompt injection is a major risk in tool-using agents because the model may obey hostile embedded instructions.
Jailbreak	Safety & Governance	An attempt to bypass model safety rules.	Coaxing the model into restricted output.	Jailbreaks target weak policy handling and can be amplified by chaining or tool access.
Data Leakage	Safety & Governance	Exposure of information that should not be revealed or used.	A model response leaking internal notes.	Leakage can happen through training, retrieval, prompts, memory, logs, or misconfigured tools.
Model Drift	Safety & Governance	Performance decay as data or behavior changes over time.	A churn model failing after pricing changes.	Monitoring and retraining are key because static deployment rarely stays optimal.
Concept Drift	Safety & Governance	The relationship between inputs and outputs changes.	Customer intent signals no longer mean the same thing.	This is often harder than data drift because the target behavior itself changed.
Data Drift	Safety & Governance	Input data distribution changes over time.	More high-mileage units entering service data.	Drift can silently erode performance even if the model code never changes.
Adversarial Attack	Safety & Governance	An input crafted to fool a model or system.	Manipulated text or image to evade detection.	Attacks may target the model itself, the retriever, the prompt, or the surrounding workflow.
Explainability	Safety & Governance	The ability to understand why a model made a decision.	Top factors driving a churn score.	Explainability improves trust, debugging, and compliance, especially on tabular models.
Interpretability	Safety & Governance	How understandable a model is to humans.	A small decision tree is highly interpretable.	Interpretability and raw predictive power often trade off, depending on the model family.
Audit Trail	Safety & Governance	A record of actions, changes, and decisions in a system.	Who approved a model release.	Audit trails are crucial for accountability in production AI.
Red Teaming	Safety & Governance	Actively testing a system for failures and abuse paths.	Trying to break an agent with hostile prompts.	Red teaming reveals issues normal happy-path testing will miss.

Relationship Table

This acts like a lightweight concept graph showing how major ideas connect.

Source Term	Relationship	Target Term	Explanation
Artificial Intelligence (AI)	includes	Machine Learning (ML)	ML is one major branch inside the broader AI field.
Machine Learning (ML)	includes	Deep Learning	Deep learning is a subtype of ML.
Machine Learning (ML)	uses	Dataset	Models learn patterns from datasets.
Dataset	contains	Feature	Features are the input variables in a dataset.
Supervised Learning	requires	Label	Supervised learning needs known target outputs.
Supervised Learning	includes	Classification	Classification is a supervised task.
Supervised Learning	includes	Regression	Regression is a supervised task.
Classification	evaluated_by	Confusion Matrix	Confusion matrices summarize classification errors.
Confusion Matrix	supports	Precision	Precision can be computed from confusion matrix counts.
Confusion Matrix	supports	Recall	Recall can be computed from confusion matrix counts.
Precision	balanced_with	Recall	Threshold choice usually trades precision against recall.
Precision	combined_in	F1 Score	F1 combines precision and recall.
Regression	evaluated_by	Mean Squared Error (MSE)	MSE is a standard regression metric.
Regression	evaluated_by	Mean Absolute Error (MAE)	MAE is a standard regression metric.
Model	trained_by	Training	Training produces learned parameters.
Training	optimizes	Loss Function	Training tries to reduce loss.
Loss Function	optimized_by	Gradient Descent	Gradient descent updates parameters using loss gradients.
Gradient Descent	controlled_by	Learning Rate	Learning rate controls step size.
Neural Network	contains	Input Layer	Networks start with input representation.
Neural Network	contains	Hidden Layer	Hidden layers learn internal representations.
Neural Network	contains	Output Layer	The output layer emits predictions.
Neural Network	learns	Weight	Weights are updated during training.
Neural Network	uses	Activation Function	Activation adds nonlinearity.
Activation Function	includes	ReLU	ReLU is a common hidden-layer activation.
Activation Function	includes	Sigmoid	Sigmoid is common for binary output.
Activation Function	includes	Softmax	Softmax is common for multiclass output.
Overfitting	mitigated_by	Regularization	Regularization reduces memorization risk.
Regularization	includes	Dropout	Dropout is one regularization method.
Regularization	includes	Weight Decay	Weight decay is another regularization method.
Unsupervised Learning	includes	Clustering	Clustering is a common unsupervised task.
Clustering	includes	K-Means	K-means is a popular clustering algorithm.
Clustering	includes	DBSCAN	DBSCAN is a density-based clustering method.
Dimensionality Reduction	includes	PCA	PCA is a standard dimensionality reduction technique.
Dimensionality Reduction	includes	UMAP	UMAP is common for visualization and clustering prep.
Reinforcement Learning	uses	Agent	An RL system centers on an agent.
Reinforcement Learning	uses	Environment	The environment responds to actions.
Reinforcement Learning	uses	State	The state summarizes the situation.
Reinforcement Learning	uses	Action	Actions are the agent's choices.
Reinforcement Learning	uses	Reward	Reward provides learning feedback.
Reinforcement Learning	learns	Policy	The policy maps states to actions.
Natural Language Processing (NLP)	includes	Tokenization	Text must be tokenized before modeling.
Tokenization	produces	Token	Tokenization breaks text into tokens.
Token	mapped_to	Embedding	Tokens are represented numerically through embeddings.
Large Language Model (LLM)	built_on	Transformer	Modern LLMs use transformer architectures.
Transformer	uses	Attention Mechanism	Attention is the core transformer innovation.
Attention Mechanism	includes	Self-Attention	Self-attention links tokens within a sequence.
Transformer	uses	Positional Encoding	Position information is needed for order.
Prompt	guided_by	System Prompt	System prompts set high-priority instructions.
Few-Shot Prompting	improves	Prompt	Examples often make prompts more reliable.
Hallucination	reduced_by	Grounding	Grounding constrains generation to evidence.
Grounding	implemented_with	Retrieval-Augmented Generation (RAG)	RAG is a common grounding pattern.
Retrieval-Augmented Generation (RAG)	uses	Retriever	A retriever finds relevant source material.
Retriever	queries	Vector Database	Semantic retrieval often uses a vector store.
Vector Database	stores	Embedding	Vector DBs index embeddings.
Hybrid Search	combines	BM25	Hybrid search often includes keyword search.
Hybrid Search	combines	Vector Database	Hybrid search also uses semantic retrieval.
Retriever	improved_by	Reranking	Rerankers refine top results.
Knowledge Base	supports	Retrieval-Augmented Generation (RAG)	RAG needs reliable source content.
AI Agent	uses	Tool Use	Agents become useful through tools.
Tool Use	implemented_with	Function Calling	Function calling is a common structured tool interface.
AI Agent	uses	Planner	Complex tasks benefit from planning.
AI Agent	uses	Executor	Plans must be executed against tools.
AI Agent	uses	Short-Term Memory	Agents rely on current working context.
AI Agent	uses	Long-Term Memory	Some agents retain information across sessions.
AI Agent	restricted_by	Guardrail	Guardrails limit unsafe behavior.
AI Agent	observed_with	Observability	Logs and traces are needed for debugging.
Human in the Loop	supports	Guardrail	Human review is a strong safety control.
Model Context Protocol (MCP)	connects	MCP Client	Clients speak MCP to access tools and resources.
Model Context Protocol (MCP)	connects	MCP Server	Servers expose capabilities over MCP.
MCP Server	exposes	MCP Tool	Servers offer callable tools.
MCP Server	exposes	MCP Resource	Servers can also expose retrievable resources.
MCP Server	exposes	MCP Prompt	Servers may offer reusable prompt templates.
MCP Tool	described_by	Tool Schema	Schemas define tool inputs and outputs.
Capability Discovery	used_by	MCP Client	Clients inspect available capabilities.
Interoperability	enabled_by	Model Context Protocol (MCP)	MCP standardization supports tool reuse across systems.
Permission Boundary	protects	MCP Tool	Permission boundaries constrain tool behavior.
Privacy	protects	Personally Identifiable Information (PII)	Privacy controls are especially important for PII.
Data Governance	supports	Privacy	Governance sets the rules for data handling.
Model Governance	supports	AI Safety	Governance manages approval and monitoring.
Prompt Injection	targets	AI Agent	Agents are vulnerable if they consume hostile content.
Jailbreak	targets	Guardrail	Jailbreak attempts try to bypass safety constraints.
Data Leakage	threatens	Privacy	Leakage can expose sensitive information.
Model Drift	degrades	Generalization	Drift lowers live performance on new data.
Data Drift	contributes_to	Model Drift	Changing inputs can trigger drift.
Concept Drift	contributes_to	Model Drift	Changing relationships can trigger drift.
Explainability	supports	Audit Trail	Explanations help audits and accountability.
Red Teaming	tests	AI Safety	Red teaming probes a system's weaknesses.

Graph JSON

Useful if you want to feed the relationships into a graph viewer, RAG index, or MCP resource later.

{
  "Artificial Intelligence (AI)": [
    {
      "relationship": "includes",
      "target": "Machine Learning (ML)",
      "explanation": "ML is one major branch inside the broader AI field."
    }
  ],
  "Machine Learning (ML)": [
    {
      "relationship": "includes",
      "target": "Deep Learning",
      "explanation": "Deep learning is a subtype of ML."
    },
    {
      "relationship": "uses",
      "target": "Dataset",
      "explanation": "Models learn patterns from datasets."
    }
  ],
  "Dataset": [
    {
      "relationship": "contains",
      "target": "Feature",
      "explanation": "Features are the input variables in a dataset."
    }
  ],
  "Supervised Learning": [
    {
      "relationship": "requires",
      "target": "Label",
      "explanation": "Supervised learning needs known target outputs."
    },
    {
      "relationship": "includes",
      "target": "Classification",
      "explanation": "Classification is a supervised task."
    },
    {
      "relationship": "includes",
      "target": "Regression",
      "explanation": "Regression is a supervised task."
    }
  ],
  "Classification": [
    {
      "relationship": "evaluated_by",
      "target": "Confusion Matrix",
      "explanation": "Confusion matrices summarize classification errors."
    }
  ],
  "Confusion Matrix": [
    {
      "relationship": "supports",
      "target": "Precision",
      "explanation": "Precision can be computed from confusion matrix counts."
    },
    {
      "relationship": "supports",
      "target": "Recall",
      "explanation": "Recall can be computed from confusion matrix counts."
    }
  ],
  "Precision": [
    {
      "relationship": "balanced_with",
      "target": "Recall",
      "explanation": "Threshold choice usually trades precision against recall."
    },
    {
      "relationship": "combined_in",
      "target": "F1 Score",
      "explanation": "F1 combines precision and recall."
    }
  ],
  "Regression": [
    {
      "relationship": "evaluated_by",
      "target": "Mean Squared Error (MSE)",
      "explanation": "MSE is a standard regression metric."
    },
    {
      "relationship": "evaluated_by",
      "target": "Mean Absolute Error (MAE)",
      "explanation": "MAE is a standard regression metric."
    }
  ],
  "Model": [
    {
      "relationship": "trained_by",
      "target": "Training",
      "explanation": "Training produces learned parameters."
    }
  ],
  "Training": [
    {
      "relationship": "optimizes",
      "target": "Loss Function",
      "explanation": "Training tries to reduce loss."
    }
  ],
  "Loss Function": [
    {
      "relationship": "optimized_by",
      "target": "Gradient Descent",
      "explanation": "Gradient descent updates parameters using loss gradients."
    }
  ],
  "Gradient Descent": [
    {
      "relationship": "controlled_by",
      "target": "Learning Rate",
      "explanation": "Learning rate controls step size."
    }
  ],
  "Neural Network": [
    {
      "relationship": "contains",
      "target": "Input Layer",
      "explanation": "Networks start with input representation."
    },
    {
      "relationship": "contains",
      "target": "Hidden Layer",
      "explanation": "Hidden layers learn internal representations."
    },
    {
      "relationship": "contains",
      "target": "Output Layer",
      "explanation": "The output layer emits predictions."
    },
    {
      "relationship": "learns",
      "target": "Weight",
      "explanation": "Weights are updated during training."
    },
    {
      "relationship": "uses",
      "target": "Activation Function",
      "explanation": "Activation adds nonlinearity."
    }
  ],
  "Activation Function": [
    {
      "relationship": "includes",
      "target": "ReLU",
      "explanation": "ReLU is a common hidden-layer activation."
    },
    {
      "relationship": "includes",
      "target": "Sigmoid",
      "explanation": "Sigmoid is common for binary output."
    },
    {
      "relationship": "includes",
      "target": "Softmax",
      "explanation": "Softmax is common for multiclass output."
    }
  ],
  "Overfitting": [
    {
      "relationship": "mitigated_by",
      "target": "Regularization",
      "explanation": "Regularization reduces memorization risk."
    }
  ],
  "Regularization": [
    {
      "relationship": "includes",
      "target": "Dropout",
      "explanation": "Dropout is one regularization method."
    },
    {
      "relationship": "includes",
      "target": "Weight Decay",
      "explanation": "Weight decay is another regularization method."
    }
  ],
  "Unsupervised Learning": [
    {
      "relationship": "includes",
      "target": "Clustering",
      "explanation": "Clustering is a common unsupervised task."
    }
  ],
  "Clustering": [
    {
      "relationship": "includes",
      "target": "K-Means",
      "explanation": "K-means is a popular clustering algorithm."
    },
    {
      "relationship": "includes",
      "target": "DBSCAN",
      "explanation": "DBSCAN is a density-based clustering method."
    }
  ],
  "Dimensionality Reduction": [
    {
      "relationship": "includes",
      "target": "PCA",
      "explanation": "PCA is a standard dimensionality reduction technique."
    },
    {
      "relationship": "includes",
      "target": "UMAP",
      "explanation": "UMAP is common for visualization and clustering prep."
    }
  ],
  "Reinforcement Learning": [
    {
      "relationship": "uses",
      "target": "Agent",
      "explanation": "An RL system centers on an agent."
    },
    {
      "relationship": "uses",
      "target": "Environment",
      "explanation": "The environment responds to actions."
    },
    {
      "relationship": "uses",
      "target": "State",
      "explanation": "The state summarizes the situation."
    },
    {
      "relationship": "uses",
      "target": "Action",
      "explanation": "Actions are the agent's choices."
    },
    {
      "relationship": "uses",
      "target": "Reward",
      "explanation": "Reward provides learning feedback."
    },
    {
      "relationship": "learns",
      "target": "Policy",
      "explanation": "The policy maps states to actions."
    }
  ],
  "Natural Language Processing (NLP)": [
    {
      "relationship": "includes",
      "target": "Tokenization",
      "explanation": "Text must be tokenized before modeling."
    }
  ],
  "Tokenization": [
    {
      "relationship": "produces",
      "target": "Token",
      "explanation": "Tokenization breaks text into tokens."
    }
  ],
  "Token": [
    {
      "relationship": "mapped_to",
      "target": "Embedding",
      "explanation": "Tokens are represented numerically through embeddings."
    }
  ],
  "Large Language Model (LLM)": [
    {
      "relationship": "built_on",
      "target": "Transformer",
      "explanation": "Modern LLMs use transformer architectures."
    }
  ],
  "Transformer": [
    {
      "relationship": "uses",
      "target": "Attention Mechanism",
      "explanation": "Attention is the core transformer innovation."
    },
    {
      "relationship": "uses",
      "target": "Positional Encoding",
      "explanation": "Position information is needed for order."
    }
  ],
  "Attention Mechanism": [
    {
      "relationship": "includes",
      "target": "Self-Attention",
      "explanation": "Self-attention links tokens within a sequence."
    }
  ],
  "Prompt": [
    {
      "relationship": "guided_by",
      "target": "System Prompt",
      "explanation": "System prompts set high-priority instructions."
    }
  ],
  "Few-Shot Prompting": [
    {
      "relationship": "improves",
      "target": "Prompt",
      "explanation": "Examples often make prompts more reliable."
    }
  ],
  "Hallucination": [
    {
      "relationship": "reduced_by",
      "target": "Grounding",
      "explanation": "Grounding constrains generation to evidence."
    }
  ],
  "Grounding": [
    {
      "relationship": "implemented_with",
      "target": "Retrieval-Augmented Generation (RAG)",
      "explanation": "RAG is a common grounding pattern."
    }
  ],
  "Retrieval-Augmented Generation (RAG)": [
    {
      "relationship": "uses",
      "target": "Retriever",
      "explanation": "A retriever finds relevant source material."
    }
  ],
  "Retriever": [
    {
      "relationship": "queries",
      "target": "Vector Database",
      "explanation": "Semantic retrieval often uses a vector store."
    },
    {
      "relationship": "improved_by",
      "target": "Reranking",
      "explanation": "Rerankers refine top results."
    }
  ],
  "Vector Database": [
    {
      "relationship": "stores",
      "target": "Embedding",
      "explanation": "Vector DBs index embeddings."
    }
  ],
  "Hybrid Search": [
    {
      "relationship": "combines",
      "target": "BM25",
      "explanation": "Hybrid search often includes keyword search."
    },
    {
      "relationship": "combines",
      "target": "Vector Database",
      "explanation": "Hybrid search also uses semantic retrieval."
    }
  ],
  "Knowledge Base": [
    {
      "relationship": "supports",
      "target": "Retrieval-Augmented Generation (RAG)",
      "explanation": "RAG needs reliable source content."
    }
  ],
  "AI Agent": [
    {
      "relationship": "uses",
      "target": "Tool Use",
      "explanation": "Agents become useful through tools."
    },
    {
      "relationship": "uses",
      "target": "Planner",
      "explanation": "Complex tasks benefit from planning."
    },
    {
      "relationship": "uses",
      "target": "Executor",
      "explanation": "Plans must be executed against tools."
    },
    {
      "relationship": "uses",
      "target": "Short-Term Memory",
      "explanation": "Agents rely on current working context."
    },
    {
      "relationship": "uses",
      "target": "Long-Term Memory",
      "explanation": "Some agents retain information across sessions."
    },
    {
      "relationship": "restricted_by",
      "target": "Guardrail",
      "explanation": "Guardrails limit unsafe behavior."
    },
    {
      "relationship": "observed_with",
      "target": "Observability",
      "explanation": "Logs and traces are needed for debugging."
    }
  ],
  "Tool Use": [
    {
      "relationship": "implemented_with",
      "target": "Function Calling",
      "explanation": "Function calling is a common structured tool interface."
    }
  ],
  "Human in the Loop": [
    {
      "relationship": "supports",
      "target": "Guardrail",
      "explanation": "Human review is a strong safety control."
    }
  ],
  "Model Context Protocol (MCP)": [
    {
      "relationship": "connects",
      "target": "MCP Client",
      "explanation": "Clients speak MCP to access tools and resources."
    },
    {
      "relationship": "connects",
      "target": "MCP Server",
      "explanation": "Servers expose capabilities over MCP."
    }
  ],
  "MCP Server": [
    {
      "relationship": "exposes",
      "target": "MCP Tool",
      "explanation": "Servers offer callable tools."
    },
    {
      "relationship": "exposes",
      "target": "MCP Resource",
      "explanation": "Servers can also expose retrievable resources."
    },
    {
      "relationship": "exposes",
      "target": "MCP Prompt",
      "explanation": "Servers may offer reusable prompt templates."
    }
  ],
  "MCP Tool": [
    {
      "relationship": "described_by",
      "target": "Tool Schema",
      "explanation": "Schemas define tool inputs and outputs."
    }
  ],
  "Capability Discovery": [
    {
      "relationship": "used_by",
      "target": "MCP Client",
      "explanation": "Clients inspect available capabilities."
    }
  ],
  "Interoperability": [
    {
      "relationship": "enabled_by",
      "target": "Model Context Protocol (MCP)",
      "explanation": "MCP standardization supports tool reuse across systems."
    }
  ],
  "Permission Boundary": [
    {
      "relationship": "protects",
      "target": "MCP Tool",
      "explanation": "Permission boundaries constrain tool behavior."
    }
  ],
  "Privacy": [
    {
      "relationship": "protects",
      "target": "Personally Identifiable Information (PII)",
      "explanation": "Privacy controls are especially important for PII."
    }
  ],
  "Data Governance": [
    {
      "relationship": "supports",
      "target": "Privacy",
      "explanation": "Governance sets the rules for data handling."
    }
  ],
  "Model Governance": [
    {
      "relationship": "supports",
      "target": "AI Safety",
      "explanation": "Governance manages approval and monitoring."
    }
  ],
  "Prompt Injection": [
    {
      "relationship": "targets",
      "target": "AI Agent",
      "explanation": "Agents are vulnerable if they consume hostile content."
    }
  ],
  "Jailbreak": [
    {
      "relationship": "targets",
      "target": "Guardrail",
      "explanation": "Jailbreak attempts try to bypass safety constraints."
    }
  ],
  "Data Leakage": [
    {
      "relationship": "threatens",
      "target": "Privacy",
      "explanation": "Leakage can expose sensitive information."
    }
  ],
  "Model Drift": [
    {
      "relationship": "degrades",
      "target": "Generalization",
      "explanation": "Drift lowers live performance on new data."
    }
  ],
  "Data Drift": [
    {
      "relationship": "contributes_to",
      "target": "Model Drift",
      "explanation": "Changing inputs can trigger drift."
    }
  ],
  "Concept Drift": [
    {
      "relationship": "contributes_to",
      "target": "Model Drift",
      "explanation": "Changing relationships can trigger drift."
    }
  ],
  "Explainability": [
    {
      "relationship": "supports",
      "target": "Audit Trail",
      "explanation": "Explanations help audits and accountability."
    }
  ],
  "Red Teaming": [
    {
      "relationship": "tests",
      "target": "AI Safety",
      "explanation": "Red teaming probes a system's weaknesses."
    }
  ]
}