Agent frameworks
Agent frameworks provide reusable building blocks for model calls, tools, task state, orchestration, and related agent behaviors. Their scope and abstractions vary.
Level 06 · Navigate the ecosystemA SHARED VOCABULARY
Less jargon.
More understanding.
Quick definitions, with a deeper lesson one click away.
Agent loops
An agent loop repeats a cycle of choosing a step, taking an allowed action, observing the result, and deciding what to do next.
Level 04 · Let AI do workAI agents
An AI agent is a system that uses a model to choose and perform steps toward a goal, often using tools and feedback. The term covers systems with very different levels of autonomy.
Level 04 · Let AI do workAI gateways
An AI gateway is an intermediary through which applications access model services. Depending on the implementation, it can centralize routing, usage tracking, limits, and credentials.
Level 06 · Navigate the ecosystemAPIs
An application programming interface, or API, is a defined way for software systems to request information or operations from each other.
Level 06 · Navigate the ecosystemApproval gates
An approval gate pauses a proposed operation until an authorized person approves the specific action under the required conditions.
Level 05 · Make it reliableChunking
Chunking splits larger source material into smaller units for indexing and retrieval. A useful chunk keeps a coherent piece of meaning together.
Level 02 · Give AI knowledgeConfidence thresholds
A confidence threshold determines when a system accepts a result and when it asks for clarification, review, or a different method. Its usefulness depends on the signal behind the score.
Level 03 · Shape the behaviorContext windows
The context window is the model’s capacity for information within one request. Instructions, conversation history, retrieved documents, tool results, and output all compete for a finite budget, subject to the model’s limits.
Level 01 · Understand AICost and latency
Cost is the resource or monetary expense of a task. Latency is the time a user waits for a useful result. Both depend on the whole workflow, not only one model call.
Level 05 · Make it reliableDeterministic vs probabilistic systems
A deterministic rule produces the same result from the same inputs under the same conditions. A probabilistic model works with likelihoods, and generation may also involve sampling from those likelihoods.
Level 01 · Understand AIEmbedding models vs LLMs
An embedding model produces representations useful for comparison. A generative large language model produces sequences, usually text. Both learn patterns, but their outputs and jobs differ.
Level 01 · Understand AIEmbeddings
An embedding is a list of numbers representing features of an input, such as text. For semantic search, the model is trained so useful similarities are reflected in the relationships between these lists.
Level 01 · Understand AIEvals
Evals are systematic tests of how well an AI system performs its intended task. They turn a vague impression of quality into evidence you can compare.
Level 05 · Make it reliableFallbacks
A fallback is a planned alternative when the preferred AI path cannot complete a task reliably or safely.
Level 03 · Shape the behaviorFew-shot prompting
Few-shot prompting supplies a small set of examples showing the desired input-to-output pattern within the request.
Level 03 · Shape the behaviorFine-tuning
Fine-tuning is additional training that adjusts a model using a selected dataset to improve a target behavior or task.
Level 03 · Shape the behaviorGPUs and inference
GPUs are processors designed for large amounts of parallel computation. They are widely used for the numerical operations involved in training and running AI models.
Level 06 · Navigate the ecosystemGrounding
Grounding means anchoring an answer in relevant evidence supplied to the system, such as a policy, a database result, or an observed tool response.
Level 02 · Give AI knowledgeGuardrails
Guardrails are checks and controls that limit unacceptable inputs, outputs, or actions around an AI system.
Level 05 · Make it reliableHallucinations
A hallucination is an output that is fabricated, unsupported, or inconsistent with the relevant evidence. It can sound just as confident as a correct answer.
Level 01 · Understand AIHandling cross-references
A cross-reference points from one part of a source to another, such as ‘subject to section 8.’ Retrieval must preserve or follow that relationship when the referenced material affects the answer.
Level 02 · Give AI knowledgeHugging Face
Hugging Face provides a platform and tools for working with machine-learning models and datasets. Its Hub hosts repositories where model artifacts and documentation can be shared.
Level 06 · Navigate the ecosystemHuman-in-the-loop
Human-in-the-loop means a person participates at a defined point in an AI-assisted process to review, correct, approve, or resolve uncertainty.
Level 04 · Let AI do workInference vs training
Training changes a model’s learned parameters using data. Inference runs an existing model on new input to produce an output.
Level 01 · Understand AILocal models
A local model runs on hardware you operate, such as a laptop, workstation, or private server, rather than sending inference requests to an external model service.
Level 06 · Navigate the ecosystemMemory
Application memory is information retained for use beyond the current interaction, such as preferences, summaries, or selected prior events.
Level 04 · Let AI do workMetadata
Metadata is descriptive information attached to content: its source, owner, date, region, document type, section, or access classification.
Level 02 · Give AI knowledgeModel Context Protocol (MCP)
MCP is a protocol for connecting AI applications with external capabilities and context through a common interface.
Level 06 · Navigate the ecosystemModel families and sizes
A model family is a related group of models. Variants may differ in size, training, supported inputs, reasoning behavior, and operational requirements.
Level 01 · Understand AIModel routing
Model routing selects a model or processing path for a task instead of sending every request to the same model.
Level 03 · Shape the behaviorModel serving
Model serving is the operational layer that accepts requests, runs a model, and returns results to applications.
Level 06 · Navigate the ecosystemNVIDIA Cosmos
NVIDIA Cosmos is a platform and family of world-model capabilities aimed at physical AI: systems that interpret, generate, or reason about physical environments.
Level 06 · Navigate the ecosystemObservability
Observability is the ability to understand a system’s behavior from signals such as logs, metrics, traces, and user feedback.
Level 05 · Make it reliableOllama
Ollama is software for using supported models through local applications and APIs. Its ecosystem includes both locally run and cloud model options.
Level 06 · Navigate the ecosystemOpen vs closed models
Models vary in how much is available to inspect, modify, and run. Open weights means the learned parameters are available; it does not necessarily mean training data, training code, or unrestricted usage rights are available.
Level 01 · Understand AIOrchestration
Orchestration coordinates the models, tools, people, and services that make up a larger process.
Level 04 · Let AI do workParent-child retrieval
Parent-child retrieval searches smaller passages but returns a larger containing section when more context is needed.
Level 02 · Give AI knowledgePermissions
Permissions define which identities may access particular data or perform particular operations. They should be enforced by the application and the systems it calls.
Level 05 · Make it reliablePrompt injection
Prompt injection happens when untrusted content tries to redirect a model’s behavior by presenting itself as an instruction.
Level 05 · Make it reliablePrompt templates
A prompt template is a reusable structure that combines stable instructions with task-specific inputs.
Level 03 · Shape the behaviorReliability vs autonomy
Autonomy is how much a system can decide and do without intervention. Reliability is how consistently it achieves acceptable outcomes. They are separate properties.
Level 05 · Make it reliableReranking
Reranking is a second pass that reorders retrieved candidates using a more focused assessment of their relevance to the question.
Level 02 · Give AI knowledgeRetrieval-augmented generation (RAG)
RAG combines finding relevant information with generating an answer from it. The system retrieves evidence at the time of the question instead of relying only on the model’s training.
Level 02 · Give AI knowledgeSemantic chunking
Semantic chunking chooses boundaries using changes in meaning or topic, rather than only a fixed character or token count.
Level 02 · Give AI knowledgeSimilarity search
Similarity search ranks items by how close their representations are to a query. In semantic search, that closeness is intended to capture useful relationships in meaning.
Level 02 · Give AI knowledgeState
State is the explicit record of where a task currently stands: inputs collected, steps completed, approvals received, actions taken, and work still pending.
Level 04 · Let AI do workStructured output
Structured output returns information in a defined format, such as a record with fixed fields and allowed categories, rather than only free-form prose.
Level 03 · Shape the behaviorSystem prompts
A system prompt is an instruction supplied by the application to define the model’s role, priorities, and expected behavior. Products differ in how they name and prioritize instruction layers.
Level 03 · Shape the behaviorTemperature
Temperature is a generation setting that changes how strongly a model favors likely next tokens. Lower settings generally concentrate the choices; higher settings allow more variation.
Level 01 · Understand AITokens
Tokens are the small units a language model processes. A token can be a word, part of a word, a number fragment, or punctuation. The split depends on the model’s tokenizer and the language.
Level 01 · Understand AITool calling
Tool calling lets a model request a defined operation from another system, such as looking up an order or preparing a calendar event.
Level 04 · Let AI do workTracing
A trace connects the steps of a single task so you can follow its path through retrieval, model calls, tools, validations, and approvals.
Level 05 · Make it reliableVector databases
A vector database stores numerical representations and supports searching for nearby vectors. Many also store text and metadata or combine vector search with other search methods.
Level 02 · Give AI knowledgeWorkflows vs agents
A workflow follows a predefined process. An agent uses a model to decide more of the process dynamically. Many practical systems combine these patterns.
Level 04 · Let AI do workNo matching concepts. Try a broader term.