Canonical Concepts
Definitions for the concepts used across CogniSwitch guides. Explore how the ideas connect, then follow the related coverage for practical explanations and examples.
Established terms use external references where an identity match exists. Operational definitions describe how a term is used in our guides. Related concepts and coverage provide the surrounding context.
AI guardrails constrain an AI application input, output, or action according to defined conditions. Different implementations detect problems, block actions, or validate a decision against explicit policy.
https://cogniswitch.ai/concepts#ai-guardrails
In CogniSwitch usage, a context graph connects knowledge with the policies, relationships, and source context needed for a decision. It is a governed knowledge graph intended to support source-backed retrieval and reasoning.
https://cogniswitch.ai/concepts#context-graph
A context window is the amount of input and output a language model can process in one interaction, measured in tokens. A larger window does not by itself resolve the meaning, authority, or consistency of the information supplied.
https://cogniswitch.ai/concepts#context-window
ContextOps is the operational work of keeping knowledge and its meaning usable as AI systems change. It covers the creation, review, maintenance, and delivery of context, with responsibilities comparable to those of other engineering operations disciplines.
https://cogniswitch.ai/concepts#contextops
Data provenance records where information came from and how it was transformed. In an AI workflow, it supports tracing an answer or decision back through its evidence and processing steps.
https://cogniswitch.ai/concepts#provenance
A database schema describes how data is organized in a database, including structures and their constraints. It specifies the storage model, which serves a different purpose from an ontology describing domain meaning.
https://cogniswitch.ai/concepts#database-schema
DataOps applies operational discipline to the delivery and maintenance of data. It coordinates people, processes, and automation so that data remains usable as sources and requirements change.
https://cogniswitch.ai/concepts#dataops
Deterministic reasoning applies explicit rules or logic so that the same inputs, rules, and system conditions produce the same conclusion. In a knowledge foundation, those reasoning steps can be inspected alongside their sources.
https://cogniswitch.ai/concepts#deterministic-reasoning
Deterministic retrieval follows explicit retrieval logic over structured knowledge under defined system conditions. It checks for the presence of relevant facts rather than relying only on text similarity.
https://cogniswitch.ai/concepts#deterministic-retrieval
DevOps brings software development and operations together to improve how software is built, released, and maintained. Shared responsibility and repeatable delivery practices connect the two functions.
https://cogniswitch.ai/concepts#devops
Explainability concerns making a system output understandable to a person. An explanation is useful context, while traceability and verification address the separate question of whether the output follows from its evidence.
https://cogniswitch.ai/concepts#explainability
A graph database stores and queries data using nodes and relationships. It can store a knowledge graph, but the database technology alone does not supply the domain meaning, governance, or source context of that graph.
https://cogniswitch.ai/concepts#graph-database
Graph RAG uses graph structures when retrieving information for a language model. The graph can make connected context available, while the reliability of the resulting answer still depends on the retrieval and generation design.
https://cogniswitch.ai/concepts#graph-rag
Grounding connects an AI output or decision to the information it is intended to rely on. In CogniSwitch usage, the relevant evidence and policies must remain identifiable so that the connection can be checked.
https://cogniswitch.ai/concepts#grounding
An AI hallucination is an output presented as factual that is incorrect or unsupported. Retrieval can provide useful evidence, but supplying evidence does not by itself guarantee that the generated answer will use it correctly.
https://cogniswitch.ai/concepts#hallucination
In the ContextOps discussion with Karthik Soman, a statement hangs as a fact when its relationships and context support treating it as one. The phrase focuses attention on the evidence connecting a statement to the rest of the knowledge, rather than on fluent wording alone.
https://cogniswitch.ai/concepts#hangs-as-a-fact
Healthcare AI applies artificial-intelligence methods to clinical, administrative, or operational healthcare work. Source-backed knowledge, policy validation, and traceability matter when these systems support regulated decisions.
https://cogniswitch.ai/concepts#healthcare-ai
Human-in-the-loop describes a workflow in which people participate in reviewing, correcting, or deciding a system output. Its effectiveness depends on what reviewers can see, how much time they have, and whether they can act on a problem.
https://cogniswitch.ai/concepts#human-in-the-loop
A knowledge graph represents entities or concepts and the relationships between them. It can connect facts to their sources and use an ontology to make the meaning of those relationships explicit.
https://cogniswitch.ai/concepts#knowledge-graph
A large language model is a language model trained on large amounts of text to learn statistical patterns used in generating and processing language. Its output alone does not establish that a statement is supported by an authoritative source.
https://cogniswitch.ai/concepts#large-language-model
LLM evaluation and observability tools measure and inspect language-model applications and their outputs. They support testing and monitoring; those measurements are distinct from enforcing an explicit policy at each decision.
https://cogniswitch.ai/concepts#llm-eval-observability
LLM-as-a-Judge uses a language model to evaluate another output against instructions or criteria. Its judgments are model-based assessments, distinct from deterministic verification against explicit rules.
https://cogniswitch.ai/concepts#llm-as-a-judge
MLOps manages the lifecycle of machine-learning systems, from development and deployment to monitoring and maintenance. It addresses the operational responsibilities that continue after a model is trained.
https://cogniswitch.ai/concepts#mlops
Neuro-symbolic AI combines neural methods, such as language models, with explicit knowledge and symbolic reasoning. Knowledge graphs and rules can supply structure that a neural model alone does not provide.
https://cogniswitch.ai/concepts#neuro-symbolic-ai
An ontology defines the concepts, relationships, and constraints used to describe a domain. It gives people and software a shared meaning for information, beyond organizing it into categories.
https://cogniswitch.ai/concepts#ontology
Phantom Human-in-the-Loop is the CogniSwitch term for oversight that exists formally but cannot reliably detect or correct failures. A reviewer may be present while volume, missing context, or limited authority makes the review ineffective.
https://cogniswitch.ai/concepts#phantom-human-in-the-loop
Policy encoding represents policies, procedures, or evaluation criteria in a form software can apply. In a knowledge foundation, it connects domain knowledge with the explicit conditions used to validate decisions.
https://cogniswitch.ai/concepts#policy-encoding
Prior authorization is a process through which a health insurer reviews whether a proposed service or treatment meets its requirements before approving coverage. Applying these requirements consistently depends on the relevant healthcare policies and evidence.
https://cogniswitch.ai/concepts#prior-authorization
A property graph is a graph data model whose nodes and relationships can carry properties. It is one way to store connected data, distinct from RDF and from the domain meaning supplied by an ontology.
https://cogniswitch.ai/concepts#property-graph
The Resource Description Framework is a W3C data model for representing statements as subject, predicate, and object. Shared identifiers allow statements from different sources to describe the same resources.
https://cogniswitch.ai/concepts#resource-description-framework
Retrieval-augmented generation retrieves external information and supplies it to a language model when producing an answer. The quality of retrieval and the reliability of generation are separate concerns.
https://cogniswitch.ai/concepts#retrieval-augmented-generation
Schema.org provides a shared vocabulary for describing entities and content in structured data on the web. Correct use of its types and properties makes the intended meaning of that markup clearer to consumers.
https://cogniswitch.ai/concepts#schema-org
A semantic layer provides shared business meanings over underlying data structures. Its usefulness depends on which definitions, relationships, and rules it makes explicit for its consumers.
https://cogniswitch.ai/concepts#semantic-layer
A semantic triple is a subject, predicate, and object used to express a statement in RDF. The predicate identifies the relationship between the subject and the object.
https://cogniswitch.ai/concepts#semantic-triple
The Semantic Web uses shared identifiers and vocabularies to publish data with machine-readable meaning. RDF, ontologies, and linked data are parts of that approach.
https://cogniswitch.ai/concepts#semantic-web
The Shapes Constraint Language is a W3C language for validating RDF graphs against declared conditions. It can check whether graph data satisfies a defined structure and other constraints.
https://cogniswitch.ai/concepts#shacl
The Simple Knowledge Organization System is a W3C vocabulary for publishing controlled vocabularies such as thesauri and taxonomies. It provides concept labels, definitions, and broader, narrower, and related relationships.
https://cogniswitch.ai/concepts#skos
SPARQL is a W3C query language and protocol for working with RDF data. It can retrieve information by matching patterns across graph statements.
https://cogniswitch.ai/concepts#sparql
A taxonomy organizes concepts or things into categories, often with broader and narrower relationships. An ontology can add other relationship types and constraints beyond that classification structure.
https://cogniswitch.ai/concepts#taxonomy
In CogniSwitch usage, traceability means following an AI output back through the source knowledge, concepts, and decision steps that produced it. It uses provenance to make that evidence chain inspectable.
https://cogniswitch.ai/concepts#traceability
Related external concept: Software traceability
The Web Ontology Language is a W3C family of languages for expressing ontologies with formal semantics. It can describe classes, relationships, and logical constraints used in automated reasoning.
https://cogniswitch.ai/concepts#web-ontology-language