Glossary

The words below are used the way the console uses them. Where a word means something narrower in Olympusx AI than in general use, the definition says so.

Agent. A model given instructions, a routing key, knowledge, tools, skills and collaborators, that runs step by step within ceilings. See Agents, tools and skills.

API key. The credential an application presents to the gateway. It belongs to one project, starts with olp_, is shown once, and can expire, be limited to named models, and be revoked. See Keys, budgets and metering.

Approval. A tool call held for a person to approve or reject before it runs. It expires unrun after 24 hours. See Human approval.

Assistant. An agent published for people to talk to: the tenant’s own people in the console, or anyone with its share link. See Assistants.

Audit log, audit trail. The record of every change and decision in a tenant and on the platform. Its entries are hash-chained, so removing or altering one breaks the chain. See Audit.

Bastion. A host you run outside the cluster that fetches model weights and runtime images for an air-gapped platform, and through which the console reaches a registered cluster’s network. Olympusx AI calls it; it never calls Olympusx AI. See Remote access.

Budget. A cap on tokens spent per day or per month, on a tenant, project, key or model. It blocks or only alerts. See Keys and budgets.

Capacity, capacity envelope. The CPU cores, memory, storage and GPUs a tenant may use on one cluster, granted by a platform administrator and divided into project quotas. See Tenants.

Catalog. The models the platform knows, shared by every tenant: open-weight models whose weights are in your registry, and external models a provider serves. See The model catalog.

Ceiling. The most tool calls, model calls or tokens one agent run may use. A run that reaches one stops.

Cluster. A Kubernetes cluster registered with Olympusx AI, where model servers and vector databases run. See Clusters.

Code source. A connection to GitLab, GitHub or Bitbucket that names groups and repositories for the Code Assistant to index. See Code sources.

Collaborator. Another agent an agent may hand a question to. See Collaborators.

Collection. A searchable set of documents in one vector database, embedded by one embedding model. See Collections.

Conversation. What a person and an assistant or agent said to each other, read as one thing. In the API, a thread.

Deployment. One model served for one tenant, in one project namespace on one cluster. See Deploy a model.

Developer key. An API key a developer mints for their own editor on the Code Assistant page. It also reaches the codebase tools. See Developer keys.

End user. The tenant role that chats with published assistants. In the identity provider, MEMBER. Also, loosely, a person who uses an assistant through its share link, who has no role and no account.

End-user reference. The opaque handle an application sends for the person a call or conversation is for (user, endUserRef or u), so that person’s records can be erased. Never personal data itself.

Erasure. Deleting one person’s conversations and captured turns on request, with a receipt. See Erasure.

External model. A model a cloud provider serves, registered in the catalog and called through the gateway. Its prompts leave your network. See External models.

Finding. In Security, a problem an assessment or investigation reports, with its evidence. In guardrails, what a detector found in a prompt or answer.

Gateway. The OpenAI-compatible entrance for API traffic, which checks every request before it reaches a model. See The gateway.

Grant. Making a model reachable by one tenant’s keys. A platform administrator grants; on a tenant’s page this is its subscription. See Models and deployments.

Guardrails. The checks on prompts and answers for personal data, secrets, prompt injection, toxicity and banned topics, set per tenant. See Guardrails.

Hunt. A Security check that runs on a schedule over the audit log, looking for one pattern. See Hunts.

Investigation. A Security question answered by reading the audit log or a log source in rounds of queries, ending in a verdict and findings. See Investigations.

LLMaaS. A model the platform offers to tenants, who subscribe to it and call it with their keys, metered in tokens and priced from the price book. Nothing is offered by default.

Log source. A connection to a tenant’s OpenSearch or Loki that Security investigations query. See Log sources.

Model budget. A tenant’s budget for one granted model, in weighted tokens, set by a platform administrator.

Offer. Making a catalog model available for tenants to subscribe to, with a token weight and a price. See LLMaaS.

Pipeline. What loads documents into a collection, from uploads or a connector, on demand or on a schedule. See Pipelines and connectors.

Platform admin, platform administrator. A person who runs the platform: clusters, model serving, tenants, billing and settings. Not a tenant role.

Price book. The prices the platform charges per model and per GPU, in the platform currency. See Billing.

Project. A division of a tenant’s work. It holds API keys, budgets and a quota on each cluster, and has a namespace on each cluster where it has a quota. See Projects.

Relevance floor. The score below which a retrieved passage is not used to ground an answer. See Knowledge and grounding.

Reranker. A model that scores passages against a question so the best are kept.

Review capture. Keeping 1 in N gateway requests for people to judge in the review queue. Off until a tenant turns it on.

Review queue. The answers no one has judged yet. See Review queue.

Role. What a person may do in their tenant: Admin, Builder, End user, Reviewer or Auditor. See Roles.

Routing key. The short description of when to use an agent, which other agents read to decide whether to hand it a question. Labelled When it is used in the agent editor.

Run. One execution of an agent, from a message to an answer or a stop, with every step it took. See Runs.

Runtime. A serving image the platform uses to run models. Platform administrators choose which are offered and which is recommended. See Runtimes.

Score, judgment. One person’s verdict on one answer: Acceptable, Wrong, Unsafe or Off policy. See Scores.

Serving pool. The GPU nodes a platform administrator put into service for models. Olympusx AI discovers GPU nodes and puts none in the pool on its own.

Share link. An assistant’s public address with its key after #k=, which lets anyone holding it talk to the assistant without signing in. See Publish and share.

Skill. A named procedure an agent loads for a kind of task, which can narrow the tools it may call. See Skills.

Slug. A short, lowercase identifier for a tenant or an assistant, used in addresses.

Tenant. One isolated customer of the platform, with its own people, projects, models, knowledge, keys, records and bill. See Tenants, projects and people.

Token weight. The multiplier that turns a model’s tokens into weighted tokens for model budgets, set from its GPU footprint.

Tool, tool server. An action an agent can take, served by a tool server the tenant registers over the Model Context Protocol. See Tools.

Trace. The record of what was said at each step of a request, with timings and tokens. See Traces.

Vector database. A tenant’s store for its collections. See Vector databases.

Workspace. The half of the console a tenant’s people use, as opposed to Platform.