165 AI, LLM, generative AI and agent-engineering terms, from foundations to emerging developer jargon.
Combine topic filters with search. Press Esc to clear search. Each new term links to its primary reference.
Artificial Intelligence is the broad field of creating systems that perform tasks normally requiring human intelligence—such as pattern recognition, language understanding, or decision-making.
First term; sets the stage.
Generative AI is a practical branch of AI focused on creating new content (text, images, code, etc.).
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 4
A subset of AI where algorithms improve at a task through experience with data instead of explicit rule-based programming.
Machine Learning is how most modern AI - including generative systems - actually learns.
Generative AI models are built using Machine Learning techniques.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 4
A layered collection of mathematical “neurons” that learn to map inputs to outputs by adjusting weights during training—loosely inspired by the brain.
Neural Networks are the dominant Machine Learning architecture powering today’s AI.
All mainstream generative models (e.g., GPT, DALL·E) are large neural networks.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 4
A Neural Network with many layers (“deep”) that can automatically learn complex patterns in data.
Deep Learning extends Neural Networks to substantial depth, enabling breakthroughs in language and vision.
Generative models rely on deep architectures to capture the nuances needed to create fluent text or realistic images.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 4
A large, curated collection of text, images, audio, or other data used to teach models about the world.
Deep Learning’s power grows with bigger, richer corpora.
The breadth and quality of a corpus determine how creative, factual, or biased a generative model may be.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 5
An individual measurable property or signal extracted from raw data to help a model learn (e.g., word position, pixel intensity).
Features are distilled from a Data Corpus before or during training.
Modern generative models often learn to create their own features internally, reducing manual engineering.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 5
The smallest unit a language model processes—commonly a word, sub-word, or even character.
Tokens are standardized features for text data.
Text generators predict one token at a time to compose sentences.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 5
The full set of tokens a model recognizes and can output.
Tokens collectively form the model’s vocabulary.
Vocabulary size affects fluency, specificity, and model size.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 5
Dense numerical vectors representing tokens, images, or other items so that semantic similarity is captured as geometric closeness.
Each token in the vocabulary maps to an embedding.
Embeddings enable models to reason about meaning and relationships, crucial for coherent generation.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 5
The multi-dimensional mathematical space in which embeddings reside; distances encode similarity.
Embeddings live and move in this Vector Space.
Manipulating vectors (e.g., arithmetic like “Paris – France + Italy ≈ Rome”) underpins many creative tasks.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 6
A neural network design that uses self-attention mechanisms to process sequences in parallel, revolutionizing language and vision tasks.
Transformers efficiently operate on sequences of embeddings in vector space.
Nearly all state-of-the-art generative language models (GPT, Gemini, Claude) are Transformers.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 6
A mechanism within Transformers that lets the model weigh the relevance of different tokens when producing each output.
Attention is the core innovation that makes Transformers powerful.
By “attending” to the right context, generative models stay coherent over long passages.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 6
The large-scale, general training phase where a model learns from a massive corpus to predict missing tokens or pixels.
Transformers are usually first pre-trained using attention on vast data.
Pre-training builds broad knowledge, enabling zero- or few-shot generation on new tasks.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 6
A very large, Pre-trained model that can be adapted to many downstream tasks.
Pre-training produces these versatile foundations.
Foundational models (e.g., GPT-4o) power chatbots, code assistants, and more.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 7
A Foundational Model specialized for text, containing billions of parameters and trained to predict the next token.
LLMs are a prominent category of foundational models.
They are the workhorses of text generation, summarization, and conversational agents.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 7
AI systems that create new content—text, images, audio, code—rather than merely analyzing existing data.
LLMs like Gemini and Claude exemplify Generative AI for language; other modalities use similar ideas.
This is the overarching topic tying all terms together.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 7
The surrounding tokens, user instructions, or real-world situation a model considers when generating an answer.
Generative AI quality hinges on effectively using context.
More context windows mean richer, more relevant outputs.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 7
A model’s ability to pick up a pattern or task on the fly from examples included in the prompt—no weight updates required.
It’s a direct outcome of how models use Context.
Enables rapid prototyping: just feed examples and the model adapts.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 7
The instruction and/or examples you provide to steer the model’s generation.
Prompts are the vehicle for In-Context Learning.
Prompt engineering is a new craft for extracting high-quality output.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 8
The phase where a trained model generates answers in response to a prompt—essentially, runtime prediction.
Inference executes the instructions embedded in a prompt.
Optimizing inference cost and speed is key for production systems.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 8
The time between submitting a prompt and receiving a model’s response.
Latency is a user-visible aspect of Inference.
Lower latency leads to smoother chat and real-time creative tools.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 8
A Graphics Processing Unit—massively parallel hardware ideal for the matrix math of training and running deep models.
GPUs help cut latency and cost during inference.
Most training and large-scale deployments rely on GPUs (or specialized accelerators).
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 8
A targeted retraining step where a Pre-trained model is adjusted on a smaller, domain-specific dataset.
Fine-Tuning may happen using GPUs after initial deployment.
Lets teams customize a foundational model to their brand voice or proprietary data.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 8
A model’s learned weight value; modern LLMs have billions.
Fine-Tuning tweaks parameters to new tasks.
Parameter count often correlates with expressiveness and memory for details.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 9
A training technique where human preferences guide the model toward more helpful or less harmful outputs.
Adjusts parameters further after Fine-Tuning.
Used to align chatbots with human expectations.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 9
The ongoing effort to ensure an AI’s goals and behaviors match human values and intentions.
RLHF is one method of pursuing Alignment.
Critical for safe, trustworthy generative systems in products.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 9
Policy, filtering, and technical controls that block disallowed or risky outputs.
Guardrails operationalize Alignment in real deployments.
Needed to prevent harmful content or data leaks when models generate freely.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 9
A sampling parameter controlling randomness during text generation: lower values make outputs more predictable; higher values more creative.
Guardrails constrain what’s allowed; Temperature shapes style within safe bounds.
A simple dial teams adjust for creativity vs. reliability.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 9
When a model confidently generates information that is false or not found in its training data.
Higher Temperature and weak Guardrails can increase Hallucinations.
Mitigating hallucinations is vital for trustworthy business applications.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 10
Quantitative or qualitative measures (e.g., accuracy, BLEU, human ratings) used to judge a model’s performance and safety.
Metrics quantify how well we control Hallucinations and meet Alignment goals.
Continuous evaluation guides improvement cycles for production-grade generative systems.
Ken Tabor · Generative AI: 30 Must-Know Terms · Source page 10
Training on examples paired with known target labels or values.
Unlike unsupervised learning, it has labeled targets.
Train an image classifier using labeled photos.
Finding structure or patterns in data without target labels.
Clustering is one unsupervised approach.
Group similar customer messages without predefined classes.
Learning a policy through interactions and reward signals.
Unlike supervised learning, feedback evaluates actions and outcomes.
Train a controller to choose actions that increase reward.
Predicting a discrete category for an input.
A supervised task distinct from numeric regression.
Classify messages as spam or not spam.
Predicting a numerical value from input features.
A supervised task distinct from classification.
Estimate a delivery time in minutes.
The examples used to fit a model's learned parameters.
Kept separate from test examples for honest evaluation.
Train a classifier on the training split.
Held-out examples used during development to choose settings or models.
Unlike the test set, it informs model selection.
Select a checkpoint using validation performance.
Held-out examples reserved to assess a model after development.
Repeated tuning on test results can contaminate the estimate.
Report final accuracy on untouched test examples.
Learning training-specific patterns that do not generalize well.
Often appears as a widening training–validation gap.
Use validation data to spot memorization.
Failing to capture useful patterns even in training data.
Contrasts with overfitting rather than merely low test accuracy.
Increase model capacity when both training and validation fit poorly.
An optimization method that updates parameters in a loss-reducing direction.
Uses gradients typically computed by backpropagation.
Update neural-network weights during training.
Computing gradients through a neural network using the chain rule.
Provides gradients to an optimizer such as gradient descent.
Calculate how each weight contributes to training loss.
A model component that converts input into internal representations.
An encoder can feed a decoder or a task-specific head.
Encode a document for classification or retrieval.
A model component that produces an output sequence from representations.
Decoder-only language models predict text autoregressively.
Generate a response token by token.
Architecture that encodes an input sequence and decodes a target sequence.
Cross-attention can connect decoder steps to encoder outputs.
Translate a sentence from one language to another.
Attention in which positions attend to representations from the same sequence.
A central component of Transformer blocks.
Relate words at different positions in a document.
Attention where queries attend to representations from another sequence or source.
Links decoder states to encoded input or other modalities.
Generate a caption conditioned on image features.
One parallel attention computation within a multi-head attention layer.
Multiple heads can learn different relationships.
Inspect attention patterns across a Transformer layer.
Position information supplied to a sequence model so token order is represented.
Complements content embeddings in attention architectures.
Distinguish swapped word orders in a sentence.
An indicator of which token positions a model may attend to.
Can hide padding or enforce causal visibility.
Prevent padded positions affecting a batched input.
Attention that prevents a position from seeing future positions.
Supports left-to-right autoregressive generation.
Train a next-token predictor without future-token leakage.
A function converting numeric scores into a normalized probability distribution.
Often turns attention scores or output logits into weights.
Convert candidate-token scores into probabilities.
A function introducing nonlinearity into a neural-network layer.
Without nonlinearities, stacked linear layers remain linear.
Apply a nonlinear transformation inside a model block.
Producing each new token conditioned on previously generated tokens.
Uses a decoding strategy to choose successive tokens.
Continue a paragraph one token at a time.
Designing and revising instructions and examples to guide model output.
Narrower than context engineering, which manages all supplied information.
Specify a format and provide representative examples.
A higher-priority instruction configuring a model's role or behavior.
Distinct from a user's request or retrieved source text.
Set assistant responsibilities before taking user input.
Requesting a task without including worked examples in the prompt.
Contrasts with few-shot prompting.
Ask a model to classify a sentence from instructions alone.
Including a small set of examples in a prompt to demonstrate a task.
Uses in-context learning without changing model weights.
Show input–output examples before a new classification.
Connecting separate model calls so earlier outputs feed later steps.
A predetermined workflow rather than an open-ended agent loop.
Generate an outline, critique it, then draft the article.
A reusable instruction structure with placeholders for variable inputs.
Keeps consistent framing while changing task-specific content.
Insert a customer question into a reviewed template.
Model output constrained or validated against a defined data shape.
Often used with typed schemas and tool interfaces.
Return machine-readable fields for downstream code.
Selecting the next token from only the k highest-scoring candidates.
A decoding restriction distinct from nucleus sampling.
Limit candidate tokens before sampling a response.
Sampling from the smallest set of tokens whose cumulative probability reaches p.
Unlike top-k, candidate-set size changes with the distribution.
Control the diversity of generated text.
Generating with relevant information retrieved from an external source.
Retrieval supplements model parameters with task-time context.
Answer a policy question from indexed documents.
Splitting source material into retrievable pieces.
Chunk boundaries affect what a retriever can return.
Divide a handbook into passages before indexing.
Repeating some content across adjacent document chunks.
Can preserve context across chunk boundaries at a storage cost.
Keep a sentence spanning two passages discoverable.
A system that stores vectors and supports similarity-oriented retrieval.
A vector index accelerates searches within such storage.
Find semantically related product descriptions.
A search structure organizing vectors for similarity queries.
May use HNSW or a flat index depending on scale.
Speed up nearest-neighbor retrieval.
Retrieving items by vector-based meaning similarity rather than exact words.
Complements keyword search in hybrid retrieval.
Match 'car repair' with documents about vehicle servicing.
Matching indexed text against the words in a query.
Useful for exact names that semantic similarity might miss.
Find a specific product code in documentation.
A lexical ranking function using term frequency and collection statistics.
Often forms the keyword branch of hybrid search.
Rank documents containing exact query terms.
Combining vector similarity results with lexical keyword results.
Balances meaning-based retrieval with exact-term matching.
Search a knowledge base containing both concepts and part numbers.
Reordering an initial set of retrieval results using an additional relevance signal.
Applied after first-stage candidate retrieval.
Promote the most relevant passages before answering.
Finding nearby vectors efficiently without guaranteeing exact exhaustive results.
HNSW is one approximate-neighbor indexing method.
Scale semantic search across millions of vectors.
A hierarchical graph index for approximate nearest-neighbor search.
Trades index complexity for fast vector queries.
Index a large embedding collection for low-latency search.
A measure comparing vector direction using the cosine of their angle.
One way to score embedding similarity.
Compare query and passage representations.
Training a pretrained model further on labeled instruction–response examples.
A distinct adaptation stage from preference optimization.
Teach a model an organization's response format.
Fine-tuning on examples of instructions paired with desired responses.
A form of supervised adaptation for following tasks.
Improve responses to varied user instructions.
Optimizing a language model from pairs of preferred and dispreferred responses.
An alternative preference-learning approach to RLHF pipelines.
Train from human pairwise response choices.
Adapting a model by training a small subset of extra or selected parameters.
Includes methods such as LoRA and prompt tuning.
Customize an LLM with lower training-memory demand.
Training low-rank weight updates while keeping original model weights frozen.
A widely used parameter-efficient fine-tuning method.
Ship a compact domain-specific adapter.
Training LoRA adapters against a quantized base model.
Combines parameter-efficient adaptation with reduced base-model memory.
Fine-tune a large model on limited GPU memory.
A small trainable component attached to a largely frozen model.
LoRA is one adapter-based approach.
Swap specialized behavior without full model copies.
Learning continuous prompt parameters while keeping the pretrained model frozen.
Unlike ordinary prompting, the prompt parameters are trained.
Learn task-specific virtual-token embeddings.
Learning a small continuous prefix to steer a frozen model.
Related to prompt tuning but modifies prefix representations.
Adapt a model to a specialized generation task.
Reusing learned representations or parameters for another task.
Fine-tuning is a common transfer-learning technique.
Adapt an image model to a smaller labeled dataset.
Training a smaller student model to imitate a teacher model.
Can transfer behavior into a cheaper model.
Build a lower-latency assistant from teacher outputs.
Representing model weights or activations with lower numerical precision.
Trades numerical precision for memory or compute savings.
Serve a model using 8-bit weights.
Stored attention keys and values reused across autoregressive decoding steps.
Avoids recomputing prior-token attention projections.
Accelerate long conversational generations.
Reusing cached attention state for requests sharing an identical prefix.
Builds on KV cache reuse across requests.
Reuse a common system prompt without recomputation.
Dynamically adding and removing inference requests from active batches.
Improves GPU utilization compared with rigid request batches.
Serve varying-length chat requests together.
An attention-memory approach that manages KV cache in blocks.
Helps reduce KV-memory fragmentation during serving.
Increase concurrent model requests on a GPU.
The inference phase that processes an input prompt before generating new tokens.
Precedes the iterative decode phase.
Measure the cost of ingesting a long prompt.
The iterative stage that emits new tokens after prompt processing.
Can reuse the attention KV cache built during prefill.
Optimize per-token serving latency.
Elapsed time from a request to its first generated token.
Includes waiting and initial prompt processing.
Track responsiveness of a streaming assistant.
Amount of work completed per unit of time, such as output tokens per second.
Different from single-request latency.
Measure aggregate serving capacity under load.
Using draft predictions that a target model verifies to speed decoding.
Targets generation speed while preserving target-model checks.
Accelerate token production with a small draft model.
The model's bounded token capacity for input and, depending on the model, output.
Context engineering chooses what occupies this capacity.
Budget documents and message history before a call.
Delivering partial model output as it is produced.
Changes perceived responsiveness without guaranteeing faster total completion.
Show answer tokens progressively in a chat UI.
A system that uses a model and tools to pursue a task through multiple steps.
An agent differs from a fixed-path workflow in who chooses steps.
Research an issue, call tools, then synthesize findings.
Repeating model decisions and tool executions until a stop condition.
Core runtime cycle underneath many tool-using agents.
Continue searching until evidence is sufficient.
A model selecting a named tool and providing its arguments.
A runtime executes the tool and returns observations to the model.
Request a database lookup through a schema.
The observation returned to an agent after a tool executes.
Becomes new context for the next model decision.
Use a search result to decide the next query.
Coordinating agent steps, tools, state, and handoffs.
Can be implemented by a harness or graph runtime.
Control a multi-step research-and-write task.
Transferring control from one agent to another specialist.
Unlike using an agent as a tool, the recipient takes over control.
Route a billing question to a billing agent.
A delegated agent working on a bounded subtask, often with separate context.
A harness can run subagents in parallel.
Delegate test-writing while a parent handles implementation.
A solution composed of multiple interacting agents.
Coordination determines how work and state are shared.
Use specialized research and verification agents.
Classifying an input and sending it down an appropriate predefined path.
An explicit workflow pattern, not necessarily an autonomous agent.
Route support tickets to different handlers.
Running independent tasks simultaneously and combining their outputs.
Trades extra concurrent work for reduced wall-clock time.
Research multiple independent sources in parallel.
A coordinator decomposes work and assigns dynamically chosen subtasks to workers.
Handles tasks whose substeps cannot all be fixed in advance.
Divide a report into specialist research assignments.
Alternating generation and feedback until a quality condition is met.
A distinct pattern from a fixed single-pass chain.
Revise a translation after evaluating terminology accuracy.
A criterion that ends an agent's iterative execution.
Can be model completion, iteration budget, or goal satisfaction.
Stop a tool loop once all required fields are verified.
Selecting and maintaining useful instructions, tools, history, and retrieved data for a model call.
Broader than optimizing the wording of a single prompt.
Trim stale tool outputs and retrieve only needed files.
Designing code, constraints, tools, tests, and feedback around a model so it can work reliably.
The harness provides an execution environment beyond model weights.
Make a coding agent run checks and inspect repository state.
Designing how agent work cycles are triggered, repeated, observed, and stopped; an emerging term.
More about execution cycles than prompt wording alone.
Run a bounded coding loop until tests pass or a limit is reached.
The surrounding runtime and configuration that gives a model tools, state, and constraints.
Harness engineering is the practice of building that runtime.
Give an agent a sandbox, planning tool, and test feedback.
Stored information describing an agent's current execution and intermediate results.
Checkpoints save state so execution can resume.
Keep plan status and accumulated research across steps.
A saved snapshot of workflow state at an execution boundary.
Enables replay, recovery, and human approval.
Resume after a failed tool rather than restarting the run.
Persisting enough progress to resume a workflow after interruption.
Uses checkpoints and careful handling of side effects.
Continue a multi-hour job after a worker restarts.
A pause that requires a person's decision before an action proceeds.
Can be implemented with graph interrupts and persistent state.
Require approval before sending an email.
Run- or thread-scoped information retained across nearby agent turns.
Usually backed by conversation state or checkpoints.
Recall the current task's previous tool results.
Information stored beyond a single run for later retrieval.
Distinct from the immediate context window.
Retrieve a past preference from a persistent store.
Condensing earlier messages or observations to free context capacity.
Trades detail for a smaller active working context.
Summarize completed tool calls before continuing.
Loading information when an agent needs it instead of placing everything in context upfront.
A context-engineering strategy that limits irrelevant tokens.
Open a referenced file only when its contents are required.
Running agent actions in an isolated, constrained environment.
A harness boundary that reduces unintended system access.
Execute generated code in a disposable workspace.
A protocol for connecting AI applications to external tools, resources, and prompts.
Separates host, client, and server roles.
Expose a documentation tool to an assistant.
The application coordinating MCP client connections and permission decisions.
A host may create a separate client for each server.
Manage tool access within a desktop AI application.
A component maintaining one connection from a host to an MCP server.
Negotiates capabilities with its server.
Connect an assistant to a repository service.
A service exposing specialized tools, resources, or prompts over MCP.
May run locally or remotely from the host.
Offer file-search functionality to multiple clients.
A callable capability advertised by an MCP server.
Tool execution is distinct from reading an MCP resource.
Invoke a repository search with typed arguments.
A server-exposed piece of contextual data that a client can read.
Resources provide context rather than executable actions.
Read a project document through a resource URI.
A server-provided reusable prompt template or message pattern.
Different from a callable MCP tool.
Load a standard code-review prompt from a server.
The initialization exchange declaring supported protocol features.
Clients and servers must respect negotiated capabilities.
Enable sampling only if both sides support it.
A structured request/response messaging protocol used by MCP.
Forms the base for messages between MCP components.
Send a typed tool-call request to a server.
Testing application behavior on a curated dataset before deployment.
Can compare changes against reference answers.
Check a new prompt on a held-out question set.
Scoring real production interactions as they occur.
Complements offline testing with live traffic coverage.
Flag format failures on sampled production traces.
A curated test collection with trusted reference cases or expected outputs.
Serves as a repeatable baseline for regression evaluation.
Keep reviewed questions and target answers for a RAG bot.
Using a language model as an evaluator under defined grading criteria.
May require reference answers and calibration with human labels.
Judge whether a summary follows provided evidence.
Comparing two candidate outputs directly rather than assigning standalone scores.
Useful when relative quality is easier to judge.
Compare two assistant responses for clarity.
Among predicted positives, the fraction that are truly positive.
Complements recall and depends on the decision threshold.
Measure how often flagged spam is actually spam.
Among actual positives, the fraction the model correctly identifies.
Can trade off against precision.
Measure how much harmful content a filter catches.
The harmonic mean of precision and recall.
Balances two classification measures in one statistic.
Compare extraction systems when both errors matter.
A repeatable check that detects behavior worsening after a change.
Can be run on a fixed evaluation dataset.
Verify a prompt edit does not break previous cases.
Instructions in input that try to redirect a model away from its intended task or rules.
Can arrive through direct messages or external content.
Treat retrieved web text as untrusted data.
Malicious instructions embedded in content an assistant retrieves or encounters.
Crosses a trust boundary when source text is treated as commands.
Ignore tool-result text requesting secret disclosure.
Giving an agent more permissions or autonomy than its task requires.
Increases the impact of failures or prompt injection.
Limit a research agent to read-only tools.
Granting only the permissions necessary for an operation.
Constrains the blast radius of agent mistakes.
Give a summarizer no delete-file capability.
A check applied to incoming content before or alongside agent processing.
Distinct from checks on final outputs and tool calls.
Reject an out-of-scope request before expensive processing.
A check validating the final answer before it is returned.
Complements input validation and tool-level checks.
Block a response containing restricted information.
A record of an operation's model calls, tool events, and handoffs.
Contains spans that help locate failures.
Inspect which tool call produced a bad answer.
A model that processes or produces more than one data modality.
Can combine text with images, sound, or video.
Answer a question about an attached chart.
Methods that process and interpret images or video.
Often provides image representations to multimodal systems.
Identify objects in a photograph.
Converting spoken audio into text.
Can provide transcribed text to a language model.
Transcribe a voice message for summarization.
Generating spoken audio from written text.
Can turn an assistant's textual answer into speech.
Read a generated explanation aloud.
A generative model that learns to reverse a noise-adding process.
A scheduler guides denoising during sample generation.
Generate an image from random noise and a prompt.
One update that moves a noisy sample toward a cleaner sample.
Repeated steps form a diffusion sampling trajectory.
Generate an image across a sequence of timesteps.
An algorithm specifying timesteps and updates for diffusion denoising.
Works with the denoising network during inference.
Choose a scheduler that trades steps for quality.
Combining conditioned and unconditioned noise predictions to steer generation.
Guidance scale controls how strongly a prompt influences output.
Adjust adherence to a text-to-image prompt.
Checks and validations placed around model or agent execution to detect, block, or handle inputs, outputs, or tool actions that violate defined rules.
An umbrella concept that includes input, output, and tool guardrails; it complements least-privilege permissions and prompt-injection defenses.
Block an unsafe request before an agent runs, or reject a final answer that fails a policy check.
Selecting which model or specialized processing path should handle a request based on task type, difficulty, cost, latency, or capability needs.
Routing can send easy requests to smaller models and harder requests to stronger models; it is related to, but narrower than, routing an entire workflow.
Classify routine support questions to a fast model while sending ambiguous cases to a more capable reasoning model.
Reusing previously computed model state for a shared prompt prefix so repeated input does not need to be processed from scratch.
It builds on reusable KV state for identical prefixes and differs from semantic caching, which can reuse whole responses for meaning-equivalent queries.
Keep stable system instructions and tool definitions at the beginning of requests so later calls can reuse the cached prefix.
The high-signal subset of available information that is actually useful for the model's current decision, rather than every token that could fit inside the context window.
A context window is the model's capacity limit; effective context is an engineering goal of filling that capacity with the smallest useful set of instructions, history, tools, and retrieved data.
Drop stale tool outputs and retrieve only the files needed for the next agent step instead of filling the full context window.
Reusing a previously generated response when a new request is semantically similar enough to a cached request, rather than requiring an exact prompt match.
Unlike prompt caching, a semantic-cache hit can skip the LLM call entirely; similarity thresholds and metadata boundaries are important to avoid returning the wrong cached answer.
Serve the same validated answer for paraphrased requests when similarity, tenant, and freshness checks pass.
An application-level allocation for how many tokens a task or model call should spend across instructions, retrieved context, conversation history, tool results, reasoning, and output.
It turns finite context and attention capacity into an explicit engineering constraint, often alongside compaction, truncation, and just-in-time retrieval.
Reserve space for the final answer and cap retrieved passages so a long tool result cannot crowd critical instructions out of the model's working context.
A curated collection of representative test cases used to measure an AI system repeatedly under defined evaluators or metrics.
It is the dataset used for offline evals; some eval sets include trusted reference answers, while others contain inputs plus rubric-based or programmatic grading criteria.
Keep a reviewed set of support questions and expected behaviors, then rerun it after prompt, model, retrieval, or tool changes to catch regressions.