Skip to main content

Glossary

Section: DOC-DX-documentation-glossary#glossary.

The terms this documentation uses, in one list. Each entry gives a one-line definition and links to the page that explains the concept in full. A term marked as replaced names the term to use instead. Read this page as Markdown at /glossary.md.

Terms​

  • Agent profile — A reusable persona with model configuration, tools and metadata describing when it should be used. Details
  • Allocation — A modeled share of compute, storage, idle capacity or shared services assigned to a project or customer; it answers a different question from provider-reported usage, an estimate or a customer charge. Details
  • Annotation queue — A named work queue assigning score configurations to human annotation items. Details
  • Callback — The request a scheduled job’s firing sends to its receiver: the configured body plus a signed identity token whose signature does not bind that body, so the receiver validates the payload against the registered schedule and acknowledges only after durable acceptance. Details
  • Client tool result — The result the caller submits for a tool call executed on the client side. Details
  • Compaction — The recorded reduction of a conversation’s retained context, described by compaction metadata. Details
  • Connected account — An external user’s OAuth connection to an app. Details
  • Conversation history — The stored messages of a conversation, from which the context management strategy selects or summarizes what the model sees. Details
  • Conversation state — The bounded read of a conversation used to retrieve a reply after an asynchronous send; it cannot correlate every queued or concurrent operation. Details
  • Conversation thread — The registry record for one conversation, kept so a user’s threads can be listed and managed together. Details
  • Dataset — An evaluation dataset: its metadata, item counts, run history and input/output schemas. Details
  • Dataset item — One evaluation input with its expected output and source trace references. Details
  • Dataset run — One execution of an evaluation dataset. Details
  • Delivery attempt — One try by the delivery backend to deliver a webhook event to an endpoint, recorded with its status and response on the event delivery. Details
  • Durable acceptance — A callback receiver’s acknowledgement of a scheduler firing only after it has stored one durable work item; return 2xx only after durable acceptance. Details
  • Durable work item — The single stored record a callback receiver creates from a scheduler firing; repeated identical deliveries produce one, and a worker completes the business action from it. Details
  • End user — A tenant’s end user: captured identity claims and editable profile values, named on backend requests by the X-On-Behalf-Of header. Details
  • Event delivery — The delivery-attempt state and response information for one webhook event. Details
  • Event provenance — The event context in a signed webhook body: the event name, id and routing metadata that identify where an emitted event came from. Details
  • Event type — The named kind of webhook event; the event name inside the signed body is the authoritative one, not a routing header. Details
  • Execution — A single firing record for a scheduled job. Details
  • Experiment context — Dataset-run attribution attached when a scripted evaluation turn is sent. Details
  • File — A stored object described by its file metadata and organised in folders. Details
  • Generation config — The canonical generation configuration applied to a request. Details
  • In flight — Accepted work that has started and not completed; a dispatch already admitted or in flight can still complete, and pausing it is not remote-effect cancellation. Details
  • Inbox acknowledgment — A user’s acknowledgement of an inbox message; it neither performs nor rolls back the business operation. Details
  • Inbox message — Notification content and interaction state displayed in a subscriber’s inbox. Details
  • MCP connection — The connection a registered custom MCP server is invoked over; refer to the MCP server itself. Use MCP server instead. Details
  • MCP server — A Model Context Protocol server an agent connects to for its available tools; a custom server is one the tenant registered and is invoked with the tenant’s credentials. Details
  • Memory — A stored statement associated with a user and, when recorded, an agent or conversation and a run or session. Details
  • Memory extraction — The pipeline that runs after a generation, derives new memory facts from the run’s messages and updates existing ones. Details
  • Memory relation — A directed edge from a source entity to a destination entity, labelled by an open relationship string. Details
  • Memory retrieval — The search that supplies stored memories to a generation from the query text submitted for it. Details
  • Message — A chat message in a conversation thread, sent by the user or generated by the assistant. Details
  • Message feedback — One user’s feedback on an assistant-generated message. Details
  • Model catalog — The catalog metadata that filters candidate models, including price, through catalog-price filters, and the context window. Details
  • Model routing — The selection of the model and provider that serve a request: explicit model ids must pass platform validation, metadata filters narrow the eligible candidates, and provider availability and run outcomes decide whether generation succeeds. Details
  • Model routing filter — Metadata-based criteria for filtering the model candidates eligible for a request. Details
  • Notification channel — One channel registered for notification delivery, with its registration state and credential count. Details
  • Observation — A timed evaluation span or generation record with model usage, inputs, outputs and scores. Details
  • On-behalf-of user — The end user a backend call made with a secret key acts for, identified by a request header; the key must carry the users:impersonate scope. Details
  • Profile version pin — A nonzero profile version that keeps the selected profile content fixed; fragments still resolve from the latest fragment set even when the profile is pinned. Details
  • Project — The scope an application’s resources and keys belong to; the current public API uses the legacy default project. Details
  • Prompt cache — The input-side cache of prompt tokens; tokens served from it are reported in the usage breakdown of prompt tokens. Details
  • Prompt fragment — Project-scoped reusable prompt-template text that profiles and other fragments include. Details
  • Provider — One of the model providers eligible for a request; provider preferences restrict and order them. Details
  • Provider preferences — Routing preferences that restrict and order the model providers eligible for a request. Details
  • Publishable key — A key configured for an accepted identity provider that lets a browser or mobile app call Travila directly with the signed-in user’s token, so the secret key never ships to the client. Details
  • Queued send — A send accepted behind an in-flight run; an accepted or queued response is not a completed reply, and the conversation can be idle without a correlated outcome for it. Details
  • Recipient — The person a notification is sent to; the requesting person must be permitted to notify that recipient, and an accepted token or supplied user id alone is not that permission. Details
  • Replay — A repeated delivery of the same scheduler firing; the receiver compares the firing identity and the payload binding before accepting it. Details
  • Response format — The structured response control that constrains the shape of the model’s output. Details
  • Response schema — The JSON schema a response format sets; when set, server validation of the model’s output defaults to enabled, with one validation retry unless configured otherwise. Details
  • Retention — Configurable retention of stored data is planned and not yet available. Details
  • Run — One accepted generation of a reply to a sent message; retaining the accepted run lets a disconnected client recover the reply without sending the question again. Details
  • Sampled coverage — The share of a fixed dataset’s eligible cases that were evaluated; not the proportion of every turn reviewed. Details
  • Schedule — The canonical representation of a scheduled job. Details
  • Scope — A label attached to an API key; where an operation enforces its scope, a missing grant returns 403 insufficient_scope. Details
  • Score — An evaluation score with its value, source, target identity and generation attribution. Details
  • Score config — A score definition specifying its value type, bounds and allowed categories. Details
  • Secret — A stored value scoped to one tenant and project, referenced by the resources that use it; read APIs return only its metadata. Details
  • Secret key — A backend API key (sk_*) that identifies the tenant and stays on the server; with the users:impersonate scope it acts for the user named in X-On-Behalf-Of. Details
  • Self-hosted — An installation of Travila your own team installs and operates, as opposed to a Travila-managed dedicated single-tenant deployment; not an Enterprise-only entitlement. Details
  • Storage — The file store behind the Storage APIs: Travila provides file storage, metadata and temporary upload and download URLs for objects your application owns. Details
  • Subscriber — The recipient of notifications, described by the contact, locale and custom attributes delivery uses. Details
  • Tool — A canonical tool or function definition, shared across all services, that the model can call. Details
  • Tool approval — The approval decision recorded for a specific tool call. Details
  • Tool call — A single structured call to a tool emitted by the model during a run. Details
  • Tool execution policy — Per-run policy for tool approvals, MCP execution concurrency and the admitted tool-call budget. Details
  • Trace — The evaluation record of one conversation turn, including its recorded conversation and message. Details
  • Turn — One position within an agentic generation run, counted against the run’s configured maximum. Details
  • Usage — Usage accounting for a generation: the tokens it consumed and the cost attributed to it. Details
  • Voice session — A Daily/Pipecat voice session and its session metadata. Details
  • Webhook endpoint — The provider record describing a webhook delivery destination and its delivery settings. Details
  • Webhook subscription — The webhook event selection and retry settings associated with an endpoint. Details
  • Workspace — In the upcoming project model, not yet available, the container for your company, with a project per application; today the public API uses the legacy default project. Details