Skip to main content

Glossary

This glossary defines the canonical terms used throughout DIAL documentation. Each term has one definition. When a term appears for the first time on a page, it should link here.

If you encounter a synonym or variant not listed below, use the canonical term instead. See the concept map for a visual overview of how these terms relate.

A​

Adapter​

A service that connects DIAL Core to an external AI provider. An adapter converts Unified API requests into the provider's API, and passes OpenAI Responses or Anthropic Messages requests through to models that support them. Adapters allow DIAL to communicate with models from providers such as Azure OpenAI, AWS Bedrock, and Google Vertex AI.

DIAL has shipped three standalone adapters, each in its own container: the OpenAI Adapter, Bedrock Adapter, and Vertex AI Adapter. The standalone adapters are replaced by DIAL Model Provider, which runs them in one service, and will be deprecated soon. Organizations can build custom adapters using the DIAL SDK.

Not to be confused with: Application (generic), Interceptor.

Addon​

Deprecated. A legacy extension mechanism based on the ChatGPT plugin protocol. The Addon concept has been abandoned and the DIAL Assistant repository that implemented it is archived. Do not use in new documentation. Modern equivalents are Application, Interceptor, and Toolset.

Agent​

An umbrella term for any conversational entity a user can interact with in DIAL — including language models and applications. Agents are listed in the Catalog and can be used as building blocks in multi-agent workflows.

See also: Agent Builder.

Agent Builder​

The canonical name for an application runner — a component that enables end users to create customized AI applications from predefined templates without writing code. Agent builders process parameters of a specific application type and expose a UI wizard for configuration.

Standard agent builders included in DIAL: Quick Apps and Code Apps.

Also known as: "application runner" (technical name in source code), "application builder" (UI label in DIAL Chat), "builder" (DIAL Admin sidebar label). These all refer to the same concept. Use agent builder in documentation.

Alias​

A deployment that forwards every call to one target: a model, an application, or a toolset. An alias has its own name, Catalog entry, and roles. A caller who may use the alias can call the target through it without rights on the target, and without reading the target's definition or credentials. Limits and cost are charged under the target's name.

Admins define aliases in the DIAL Core configuration or create them at run time in public storage. Aliases are available from DIAL Core 0.50.0. See Aliases and routers.

App Builder​

A Python-based utility that downloads source code from DIAL file storage and prepares files to build container images for Code Apps. Not to be confused with agent builder, which is the no-code UI concept.

Repository: ai-dial-app-builder-python.

Application​

A first-class extension in DIAL that DIAL Core calls as a deployment. An application serves Unified API chat completion or embedding endpoints, often built with the DIAL SDK, and can also serve the OpenAI Responses API, the Anthropic Messages API, or MCP. Applications can be invoked directly by users or composed as building blocks in multi-agent workflows.

Applications come in two flavors: schema-rich applications (defined by a JSON schema and configurable via API or UI) and applications without schemas (logic embedded in application code and container).

Not to be confused with: Adapter (translates external provider APIs), Interceptor (middleware that modifies requests/responses), or the deprecated Addon.

Application Type​

A schema-rich template that defines the structure, properties, endpoints, and optional UI wizard for a category of applications. Application types allow end users to create instances of a specific kind of application without writing code.

Standard application types included in DIAL: Quick App and Code App.

Organizations can define custom application types and register them in DIAL Core.

Assistant (deprecated)​

Deprecated. The DIAL Assistant was an implementation of the ChatGPT plugin protocol. The repository (ai-dial-assistant) is archived. Do not use in new documentation. The modern replacement is Application.

Attachment​

A file included in or produced by a chat completion request or response. Applications built with the DIAL SDK can accept and return attachments of various types. Attachments are stored in the DIAL file storage.

C​

Catalog​

The single-entry point in DIAL Chat for browsing and accessing all available models, agents, toolsets, skills, and prompts. The Catalog respects role-based access control, so each user sees only the resources available to their permissions. You can search it and filter it, for example by topic or to show only your own items.

The Catalog also serves as a collaboration hub where users can share and publish their creations.

Also known as: Marketplace (the former name).

Code App​

An application type that allows users to develop, deploy, and run Python applications directly in the DIAL Chat UI. Code Apps are useful for rapid prototyping and proof-of-concept work. They run in a secure, isolated environment managed by the DIAL platform and do not have internet access.

See also: Quick App.

Conversation​

A dialogue between a user and an agent in DIAL. Each conversation maintains its own context — messages in one conversation are not shared with another. Conversations are stored server-side and accessible from any device. Conversations can be shared, published, replayed, or played back.

Custom UI​

A fully custom user interface that replaces the standard DIAL Chat conversational UI during interactions with a specific application. Custom UIs can implement non-conversational interfaces.

Custom UIs are defined through the applicationTypeViewerUrl property in an application type schema.

D​

Deployment (configuration sense)​

A named endpoint in DIAL Core configuration that exposes a model or application to clients. A deployment defines the connection between a logical name (used in API requests) and the actual service endpoint, along with associated upstreams, interceptors, and access control settings.

Not to be confused with: Helm deployment or Kubernetes deployment (infrastructure sense). When the context is ambiguous, qualify as "DIAL deployment" or "Kubernetes deployment."

DIAL​

Deterministic Integrator of Applications and Language Models. An open-source, enterprise-grade AI platform by EPAM that provides a unified gateway to language models, application orchestration, access control, observability, and collaboration features. Always written in all-caps: DIAL, never "Dial" or "dial."

Licensed under Apache License, Version 2.0.

DIAL Admin​

The administration interface for the DIAL platform. DIAL Admin Frontend connects directly to Core for configuration, artifacts, and governance; optional backends add deployment management, evaluation, analytics, and legacy administration workflows. See Administering DIAL for the user guide.

DIAL Analytics​

The optional service that ingests DIAL Core usage logs from blob storage into ClickHouse, keeps a PostgreSQL catalog of queryable tables, and serves governed queries over them through a structured query format, SQL over REST, and optional Arrow Flight SQL. Its API is the Analytics Data Access Service (ADAS). DIAL Admin and BI tools such as Grafana query it.

Documentation: DIAL Analytics. Repository: ai-dial-analytics.

DIAL Analytics Enrichment Runner​

The optional service that executes the SQL and LLM enrichment pipelines defined in DIAL Analytics, with persistent scheduling, retries, and spend tracking. It keeps its state in its own PostgreSQL database. Aggregate rollups run inside DIAL Analytics, not in the runner.

Documentation: DIAL Analytics Enrichment Runner. Repository: ai-dial-analytics-enrichment-runner.

DIAL Chat​

The default web-based user interface for the DIAL platform. DIAL Chat provides a conversational interface, the Catalog, no-code agent builders, collaboration features, and extensibility through custom UIs and visualizers.

Repository: ai-dial-chat. Not to be confused with: DIAL Overlay (an embeddable subset of DIAL Chat).

DIAL Core​

The central Java service and only mandatory component of the DIAL platform. DIAL Core serves DIAL's AI APIs — the Unified API, OpenAI Responses, and Anthropic Messages — and provides the LLM gateway, load balancing, authentication, access control, interceptor orchestration, file storage, cost management, and observability.

DIAL Core is headless — it functions without a UI. All other DIAL components are optional.

Repository: ai-dial-core. Not to be confused with: "the backend" (too vague — use DIAL Core).

DIAL Helm​

The Helm chart repository for deploying DIAL on Kubernetes. Stable assemblies that combine compatible versions of all DIAL components are published here.

Repository: ai-dial-helm.

DIAL Log Parser​

A batch tool that converts DIAL Core request logs into Parquet datasets partitioned by deployment and date, for offline analysis and evaluation. It processes one date per run and exits.

Documentation: DIAL Log Parser. Repository: ai-dial-log-parser.

DIAL Model Provider​

The DIAL service that runs the official model adapters in one image and one process: the OpenAI, Vertex AI, Bedrock, and DIAL-to-DIAL adapters. Each adapter is served under its own path prefix, /compat/<adapter>, and keeps its own DIAL API, models, and settings. DIAL Model Provider replaces the standalone adapters.

Documentation: DIAL Model Provider. Repository: ai-dial-model-provider.

DIAL Overlay​

A library that allows embedding DIAL Chat in external web applications via an iframe event-based protocol. Overlay enables third-party applications to integrate DIAL conversational capabilities without building a custom UI from scratch.

Documentation: DIAL Chat Overlay.

DIAL Scheduler​

The optional service that runs chat completions, Responses API calls, and REST or SSE requests on recurring or one-time schedules. Each run uses its owner's delegated credentials, so it works while the user is away. It keeps schedules and run history in PostgreSQL and powers scheduled tasks in DIAL Chat.

Documentation: DIAL Scheduler. Repository: ai-dial-scheduler.

DIAL SDK​

A Python framework (Python ≥3.11) for creating applications and adapters for DIAL. Applications and adapters built with the SDK are fully compatible with the Unified API.

Repository: ai-dial-sdk. See also: Interceptors SDK.

DIAL-to-DIAL Adapter​

A development utility that allows a local DIAL instance to proxy requests to a remote DIAL Core. Useful during application development when you want to test against a remote environment without deploying locally.

Repository: ai-dial-adapter-dial. DIAL Model Provider also serves it under /compat/dial.

Dynamic Settings​

DIAL Core configuration files that can be reloaded at runtime without restarting the service. Dynamic settings define deployments, interceptors, roles, routes, API keys, and other operational parameters.

Administrators can trigger a reload via the DIAL Core API or through DIAL Admin.

E​

Embedding API​

A DIAL API endpoint that provides unified access to embedding models from any supported provider. The Embedding API normalizes provider-specific embedding interfaces into a single protocol, supporting asymmetric models, instruct models, and multi-modal embeddings.

F​

File Storage​

The persistent storage layer for DIAL, used to store conversations, prompts, applications, user files, and other resources. DIAL supports cloud blob storage (AWS S3, Google Cloud Storage, Azure Blob Storage) or a local file system. Valkey (a Redis-compatible in-memory store) is deployed on top as a cache.

DIAL does not require a centralized database — file storage and Valkey are the only persistence dependencies.

G​

Generic RAG​

An application type that answers questions over a collection of documents uploaded and indexed in advance. Each application of the type has its own documents, indexes, and pipeline settings, and is also an MCP server.

Also known as: "channel", the backend's term for one application's documents and settings. Use Generic RAG for the type, and DIAL Generic RAG for the service.

Documentation: DIAL Generic RAG.

I​

Interceptor​

Middleware that modifies requests and/or responses flowing through DIAL Core. Interceptors are inserted into deployments and execute before or after chat completion requests reach the target model or application.

Interceptors fall into three categories: pre-interceptors (modify the incoming request), post-interceptors (modify the outgoing response), and generic interceptors (modify both). Common use cases include PII detection and redaction, content filtering, prompt injection detection, and compliance enforcement.

When multiple interceptors are configured, they execute in sequence: global interceptors first, then application-type interceptors, then local interceptors. The response travels back through the chain in reverse order.

Not to be confused with: Adapter (translates external APIs), Application (business logic).

Interceptors SDK​

A Python framework for creating interceptors for DIAL chat completion and embedding models.

Repository: ai-dial-interceptors-sdk. Not to be confused with: DIAL SDK (for applications and adapters).

L​

Load Balancer​

A feature of DIAL Core that distributes requests across model upstreams. The load balancer supports weighted distribution across deployments, regions, and cloud subscriptions, and can prioritize provisioned throughput unit (PTU) deployments over pay-per-token options.

Configuration is defined through upstream parameters in dynamic settings and can be adjusted without redeployment.

M​

MCP Server​

A service conforming to the Model Context Protocol (MCP) that extends AI application capabilities with external tools and data sources. DIAL supports two integration approaches: connecting to external MCP servers hosted outside the platform, and deploying custom MCP servers as Docker containers through DIAL Admin.

MCP servers are used through toolsets in Quick Apps and other applications.

Multimodality​

The ability of DIAL applications and models to handle non-textual content alongside text. This includes image-to-text, text-to-image, text-to-video, image-to-video, and file transfers. Multimodality is supported through the Unified API and rendered in DIAL Chat through attachments and visualizers.

My Workspace​

The personal area in DIAL Chat where a user manages their bookmarked agents, toolsets, and agent builders. Resources added from the Catalog appear here. My Workspace also provides access to no-code application creation tools.

P​

Parameterized Replay​

A variant of replay where specific parts of conversation messages are replaced with variables. When the conversation is replayed or shared, users are prompted to provide their own values for those variables, creating a personalized experience from a shared template.

Per-Request Key​

A short-lived API key generated by DIAL Core for the duration of a single request. Per-request keys manage file access permissions for applications, enable distributed tracing, and attribute costs to the originating user or application. They are automatically invalidated when the request completes.

Playback​

A mode in DIAL Chat that reproduces a conversation exactly as it occurred, without re-submitting prompts to any model. Playback simulates the conversation like a recording. Not to be confused with replay, which re-submits prompts and may produce different results.

Private Space​

The logical storage area in DIAL where resources are accessible only to their owner and users with whom they have been explicitly shared. By default, all user-created resources reside in private space.

See also: Public Space.

Prompt​

A reusable text template stored in DIAL that users can invoke in conversations. Prompts can contain variables (e.g., {{country|Japan}}) with optional default values. Prompts can be shared or published.

Publication​

The workflow through which a user submits resources from their private space to the public space, making them available to all authenticated users or a restricted audience. All publication requests require administrator approval. Published applications and toolsets appear in the Catalog.

Not to be confused with: Sharing (user-to-user, no approval required).

Public Space​

The logical storage area in DIAL where resources are accessible to all authenticated users by default. Subfolders within public space can have access rules that restrict visibility to specific roles. Resources enter public space through publication or direct configuration by administrators.

See also: Private Space.

Q​

Quick App​

An application type and no-code orchestrator, conceptually similar to OpenAI's GPTs, that simplifies the creation of multi-agent workflows. Quick Apps can use agents, toolsets, REST APIs, language models, and other DIAL resources as building blocks. Configuration is defined through a JSON schema and can be managed via API or the DIAL Chat UI wizard.

Common use cases: RAG-like applications with predefined sources, applications that call external APIs via MCP toolsets, and multi-step workflows combining multiple agents.

See also: Quick App 2.0, Code App.

Quick App 2.0​

An evolution of Quick Apps that provides a composer for building applications from reusable components. Quick App 2.0 enables task-oriented workflows combining DIAL agents and external integrations (MCP servers), with any AI model that supports tool use acting as an orchestrator.

R​

Replay​

A feature in DIAL Chat that reproduces a conversation by re-submitting the original prompts, optionally with different settings such as a different model or temperature. The replayed conversation appears as a new conversation tagged [Replay]. Useful for comparing model responses to identical inputs.

Not to be confused with: Playback (exact reproduction without model calls).

Resource​

Any object managed by DIAL Core: applications, conversations, prompts, files, and toolsets. Resources can reside in private space or public space and are subject to access control rules. Resources can be shared and published.

Role​

A named set of permissions and limits assigned to JWTs or API keys in DIAL Core. Roles determine which resources a subject can access, enforce usage limits (token rate limits, cost limits), control access to system features (e.g., the admin console), and configure sharing limits.

Roles are defined in dynamic settings or through DIAL Admin.

Route​

A configuration entry in DIAL Core that proxies requests on matching paths and HTTP methods to an upstream server, such as an external API. DIAL Core handles authentication and authorization between the client and the upstream, and can balance requests across several upstreams. A route passes requests through without translating them.

Not to be confused with: Router.

Router​

A deployment that picks the model serving each request. Clients call a router by name wherever they would call a model. By default, DIAL Core forwards the request to the router service, which calls the chosen model back through DIAL Core. In decision mode, DIAL Core asks the router which deployments may serve, in order of preference, and serves the first usable one itself. A router has no limits of its own; the models it reaches are checked and charged against the caller.

Not to be confused with: Route. See Aliases and routers.

S​

Schema-Rich Application​

An application whose structure is defined by a JSON schema conforming to the DIAL Core meta schema. Schema-rich applications can be created and configured through the DIAL Core API or UI wizards, and their properties can be modified without redeploying containers.

Contrast with: applications without schemas, where business logic properties are embedded in application code and cannot be changed via API.

Sharing​

A collaboration feature that allows a resource owner to grant access to specific users via a sharing link, without publishing the resource to the entire organization. Shared resources remain in the owner's private space. The owner retains control and can revoke access at any time. Sharing links can optionally grant editing and re-sharing permissions.

Not to be confused with: Publication (organization-wide, requires admin approval).

Stage​

A structured step in an AI-generated response that shows the intermediate reasoning or actions an agent took to produce its final output. Stages are rendered in DIAL Chat as expandable sections within the conversation, providing transparency into multi-step or agentic workflows.

Stages are returned by applications through the stages field in the Unified API response.

Starter Button​

A predefined prompt button displayed at the beginning of a conversation in DIAL Chat. Starter buttons let users begin a conversation with a single click instead of typing. Configured through the starters property in a Quick App or application schema.

T​

Tool​

A capability that an AI model can invoke during a conversation to perform a specific action — such as calling an external API, querying a database, or executing a computation. Tools are declared as part of a chat completion request and conform to the OpenAI function calling convention.

In DIAL, tools can be provided by MCP servers, other applications, or models deployed in DIAL Core. See also: Toolset.

Toolset​

A named collection of tools exposed through an MCP server connection. Toolsets serve as connectors between Quick Apps (and other applications) and external services. Users can browse available toolsets in the Catalog, add them to My Workspace, and use them in their applications.

Toolsets support authentication (user credentials or organization-wide credentials) and can be shared or published.

Also written as: "toolset" (single word, used in API field names and some UI labels). In documentation, use toolset (two words) in prose.

U​

Unified API​

DIAL's extension of the OpenAI Chat Completions and Embeddings APIs, served by DIAL Core for language models, embedding models, and applications. It adds DIAL-specific features including attachments, stages, custom content rendering, state management, and interactive controls. DIAL Chat and applications built with the DIAL SDK use it.

The Unified API is one of the AI APIs that DIAL Core serves. Clients that already speak the OpenAI Responses API or the Anthropic Messages API call DIAL Core in that API instead. See AI API overview.

Not to be confused with: "DIAL API" (too vague — use Unified API when referring to the chat/embeddings protocol).

Upstream​

A backend endpoint that serves requests for a specific deployment. Each deployment can have multiple upstreams with different weights and tiers. The load balancer distributes requests across upstreams based on the configured strategy.

Upstream parameters include endpoint, key, tier, and weight.

V​

Visualizer​

A special-purpose application used to render specific content types within the DIAL Chat UI. Visualizers extend the default rendering capabilities beyond built-in Markdown and Plotly support. Custom visualizers are built using the DIAL Chat Visualizer Connector library.

Examples: rendering 3D protein structures, financial charts, or custom data formats in the chat interface.

Deprecated Terms​

TermStatusReplacement
AddonArchived. ChatGPT plugin protocol, abandoned.Application, Interceptor, Toolset
AssistantArchived. Repository ai-dial-assistant.Application
"the backend"Informal. Too vague.DIAL Core
"the frontend"Informal. Too vague.DIAL Chat
"DIAL API"Ambiguous. Could refer to any DIAL endpoint.Unified API (for the chat/embeddings protocol)

Further reading​