Glossary
Use this glossary for short canonical definitions. For the request flow and model-group ownership model, see Concepts.
| Term | Definition | Canonical page |
|---|---|---|
| allow list | The model groups a caller token may request. Requests outside the list fail before upstream credentials are used. | Available Models And Access |
| API dialect | The caller-facing API shape, such as OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages. | API Compatibility |
| attempt | One upstream provider/model try for a request. A request can have multiple attempts when fallback is configured. | Usage Reporting |
| attempt timeout | The time budget for one upstream attempt before the router can fail or move to fallback. | Router Configuration |
| cache eligibility | The request and target conditions that decide whether a response may be served from or written to cache. | Router Configuration |
| cache key | A non-secret fingerprint used to identify reusable cached responses after configured redaction and eligibility checks. | Router Configuration |
| caller | The authenticated application, service, or user represented by a router token and metadata. | Available Models And Access |
| caller token | A router-issued bearer token with allow lists, caller metadata, limits, and optional admin privileges. | Available Models And Access |
| content-admin | An operator subject authorized for governed content-capture maintenance actions such as delete or purge. | Admin Authorization |
| contract | A model-group capability and quality promise enforced before target selection. | Model Group Contracts |
| decision telemetry | Safe scalar routing evidence recorded for policy, eligibility, fallback, score, and filter decisions. | Usage Reporting |
| dialect | A provider or caller API shape used to translate requests and responses without changing the caller contract. | API Compatibility |
| effort enum | A reasoning-control value such as low, medium, or high when a provider supports effort-style reasoning selection. | Reasoning Routing |
| eligibility | The request-shape and policy checks a target must pass before it can be selected. | Routing Strategy Decision Tree |
| fallback | A retry path to another eligible target after a retryable upstream failure or configured failover order. | Customer-Controlled Routing |
| fingerprint | A safe identifier or hash used for diagnostics, cache decisions, or routing evidence without exposing raw content or tokens. | Usage Reporting |
| image input | Image-bearing request content sent through a supported VLM-capable API shape. | Image Analysis And VLM Routing |
| input modality | A target-supported input type such as text or image. | Model Metadata |
| metrics-admin | An operator subject authorized to read global /metrics; ordinary caller keys receive 403 metrics-forbidden. | Observability |
| model group | A caller-facing, deployment-defined policy name that owns a target list and routing strategy. | Concepts |
| output modality | A target-supported output type, normally text unless validated otherwise. | Model Metadata |
| owner user | The configured owner identity for caller keys, reporting, access, and usage grouping. | Router Configuration |
| policies | Authentication, authorization, routing, limits, contracts, traffic shaping, retention, and deployment rules that govern requests. | Customer-Controlled Routing |
| policy service | A trusted deployment service called by strategy: external to choose from eligible targets. | External Routing Policy Service |
| project | A deployment-owned grouping for caller access, usage reports, and operational ownership. | Router Configuration |
| project_membership | The configured relationship that authorizes a user or service identity inside a project/environment domain. | Router Configuration |
| provider catalog | Metadata about upstream providers and models, including model IDs, pricing, modalities, tool support, and validation notes. | Model Metadata |
| provider-route | A configured provider/model route exposed only through model-group targets, not directly to callers. | Add A Provider Or Model |
| reasoning control | Request fields that ask for provider reasoning or thinking behavior and become target eligibility requirements. | Reasoning Routing |
| reasoning mode | A target's validated reasoning behavior, such as supported effort levels or thinking controls. | Reasoning Routing |
| request shape feature | A safe request property such as API skin, tools, image input, reasoning, output cap, prompt size, or structured-output request. | Dynamic Score Routing |
| reservation | A quota or traffic-shaping hold based on estimated input and requested output budget before upstream completion. | Error Reference |
| route strategy | The configured mechanism that selects a target: static, failover, weighted, dynamic score, script, external, or contract-backed. | Routing Strategy Decision Tree |
| router endpoint | The deployment URL callers use instead of direct provider endpoints. | Hosted Quickstart |
| routing policy | The deployment-owned rules and strategy that choose among eligible targets inside one requested model group. | Customer-Controlled Routing |
| skin | A client compatibility surface, such as OpenAI Chat, OpenAI Responses, or Anthropic Messages. | API Compatibility |
| target | One configured provider/model entry inside a model group. | Concepts |
| target selection | The process of filtering eligible targets and choosing one according to the model group's strategy. | Routing Strategy Decision Tree |
| target-level context eligibility | A target-specific check that skips targets unable to satisfy the request context, such as max-token or modality requirements. | Router Configuration |
| tier | Deployment-defined target metadata used by policy, reports, or scripts to group targets by role or cost class. | TypeScript Routing Policy |
| token budget | The configured or requested token allowance used for quota admission, traffic shaping, and upstream output caps. | Error Reference |
| tool dialect | The tool-call protocol a target has been validated to support, such as OpenAI Chat tools or Anthropic client tools. | Agents, Tools, And Vision |
| usage rows | Relational usage database records for request, attempt, cost, latency, routing, quota, and diagnostic facts. | Usage Reporting |
| usage_db | The configured relational database used for usage reports, cost accounting, and operational triage. | Usage Reporting |
| VLM | Vision-language model behavior for image-bearing prompts, OCR, screenshots, diagrams, or browser-control context. | Image Analysis And VLM Routing |