AI engineering glossary

An A/B test randomly assigns traffic between variants to estimate a causal effect.

Accuracy is the fraction of predictions that are correct.

An activation function transforms a neuron’s pre-activation value.

An agent harness is the runtime that supplies tools, state, policy, and verification around a model.

An agent loop repeatedly observes state, chooses an action, and evaluates the result.

Agent memory preserves useful state beyond the model’s immediate context.

A2A defines communication patterns for agents that delegate and exchange task state.

An AI agent selects and executes actions toward a goal using observed state.

AI red teaming systematically searches for harmful, insecure, or unreliable model behavior.

An artificial neuron applies a weighted sum followed by an activation function.

An autoencoder learns to reconstruct an input through a constrained latent representation.

Backpropagation applies the chain rule to compute gradients through a computation graph.

Batch size is the number of examples used to estimate one gradient update.

A benchmark is a specified dataset, protocol, and metric used for comparison.

The bias-variance tradeoff balances systematic underfitting against sensitivity to training data.

BM25 is a sparse ranking function based on term frequency, document frequency, and length normalization.

Calibration measures whether predicted probabilities match observed frequencies.

A canary deployment sends a small share of production traffic to a new version before wider rollout.

A causal mask prevents a decoder token from attending to future tokens.

Chunking divides source material into units that can be indexed and retrieved.

Classification predicts one or more discrete categories.

A confusion matrix counts predictions by actual and predicted class.

A context window is the maximum token span available to one model invocation.

Continuous batching admits and removes requests as individual decoding slots become available.

A convolution slides a shared filter across local regions of an input.

Cosine similarity measures the angle between two vectors.

Data poisoning manipulates training or retrieval data to change system behavior.

Decode is the autoregressive phase that generates one new token per active sequence per iteration.

Dense retrieval ranks items using learned vector representations.

Dropout randomly masks activations during training to reduce co-adaptation.

An embedding maps a discrete item to a learned dense vector.

An epoch is one pass through the training dataset.

Experiment tracking records configurations, artifacts, and results for model runs.

F1 is the harmonic mean of precision and recall.

A feature is a measurable input supplied to a model.

A feature store manages reusable feature definitions and values for training and serving.

Fine-tuning continues training a pretrained model on a narrower objective or dataset.

Gradient descent updates parameters in the direction that reduces loss.

Grounding ties a generated claim to supplied evidence or an authoritative source.

gRPC is a typed remote procedure call system commonly carried over HTTP/2.

A guardrail is a control that constrains inputs, outputs, or actions around a model.

A hallucination is a fluent model output that is unsupported or false in the relevant context.

HTTP is a request-response application protocol for transferring representations and commands.

Human in the loop places explicit human judgment inside an automated process.

Hybrid search combines sparse lexical ranking with dense semantic ranking.

A hyperparameter is a setting that controls model structure or training.

An idempotent operation has the same intended effect when repeated with the same input.

A jailbreak is an input designed to bypass a model’s behavioral safeguards.

JSON Schema describes and validates the structure of JSON values.

Knowledge distillation trains a smaller student to match signals from a larger teacher.

A KV cache stores attention keys and values from earlier tokens during autoregressive decoding.

A label is the target value a supervised model learns to predict.

The learning rate scales each optimizer update.

LoRA learns low-rank update matrices while keeping the original model weights frozen.

A loss function converts prediction error into a scalar training objective.

MCP standardizes how model applications discover and invoke external tools and resources.

Model drift is a change over time that degrades or alters model behavior.

Model monitoring measures deployed model inputs, outputs, quality, latency, and failures over time.

A model registry records versioned model artifacts and their lifecycle metadata.

Model serving exposes a trained model through a reliable runtime interface.

A multi-agent system coordinates multiple decision-making agents around a task.

Multi-head attention runs several learned attention projections in parallel.

Next-token prediction estimates a probability distribution for the token after a prefix.

A normalization layer rescales activations using statistics over selected dimensions.

Overfitting occurs when a model learns training-specific detail that does not generalize.

A parameter is a learned numeric value inside a model.

A planning agent decomposes a goal into coordinated steps before or during execution.

Precision is the fraction of predicted positives that are truly positive.

Prefill processes all prompt tokens to initialize attention state before decoding.

Prompt injection is untrusted content that attempts to redirect a model’s instructions or tool use.

Quantization represents model values with fewer bits or a smaller numeric format.

A rate limit bounds how many operations a caller may perform in a time interval.

ReAct alternates model reasoning, actions, and observations in a task loop.

Recall is the fraction of actual positives that the system finds.

Recall@k is the fraction of relevant items found among the top k results.

A receptive field is the input region that can influence one feature value.

Regression predicts a continuous numeric quantity.

Regularization changes training to discourage brittle or overly complex solutions.

A reranker rescoring a small candidate set with a more precise relevance model.

A residual connection adds a block’s input to its transformed output.

REST is an architectural style that models operations around resources and uniform interfaces.

RAG retrieves external evidence and places it in the generation context.

Reinforcement learning from human feedback optimizes a model using preferences derived from people.

ROC AUC measures ranking quality across classification thresholds using true-positive and false-positive rates.

Self-attention mixes information among positions in the same sequence.

Server-sent events stream text events from an HTTP server to a client.

Speculative decoding uses a cheaper model to propose tokens that a target model verifies in parallel.

Supervised learning fits a mapping from inputs to known target outputs.

A supervisor agent routes work, monitors progress, and decides when to stop or escalate.

Temperature rescales logits before sampling to control distribution sharpness.

Tensor parallelism splits individual tensor operations across devices.

Time to first token measures latency from request arrival until the first generated token is available.

A token is one unit in the sequence processed by a language model.

Tokenization converts text into a sequence of vocabulary identifiers.

Tokens per second measures generation throughput over time.

Tool calling lets a model request a typed external operation.

A transformer is a sequence architecture built around attention and position-wise feed-forward blocks.

Unsupervised learning finds structure in data without supplied target labels.

A vector database stores embeddings and searches for nearby vectors.

WebSocket provides a persistent full-duplex message channel between two peers.