Skip to main content

Glossary

Clear definitions for common terms used across the Crusoe AI Platform. Each entry links to relevant guides for detailed instructions.

A​

  • Acknowledge (ack) — Telling Pub/Sub that a message has been processed successfully so it is not sent again. Unacknowledged messages stay in the queue. See publish and consume.
  • Ack deadline — The time limit allowed for processing a pulled message before the platform considers it failed and redelivers it (between 1 and 600 seconds, defaulting to 30 seconds). See topics and subscriptions.
  • Agent — A hosted AI workload built from Python agent code that provides conversational API endpoints with built-in session state, memory, and tool integration. See Agents overview.
  • API key — A secure string credential formatted as cai_... used by scripts to authenticate without entering a password. Shown only once when generated. See service accounts and API keys.
  • Apply — The action that delivers bound secret environment variables to a workload and rolls out a fresh revision. See manage secrets.
  • At-least-once — The delivery guarantee for Pub/Sub messaging. Ensures every published message is delivered, with subscribers handling potential duplicates. See publish and consume.
  • Audit log — A permanent, append-only record tracking all configuration changes, deployments, and access events in a project. See break-glass and audit.
  • Automation token — A machine access token configured for automated platform management. See API authentication.

B​

  • Backlog — The collection of published Pub/Sub messages waiting to be acknowledged by a subscription. See publish and consume.
  • Backoff — An automated retry pattern that increases the wait time between consecutive failed request attempts. See publish and consume.
  • Bearer token — An access credential sent in the HTTP Authorization header (Authorization: Bearer <token>) to authenticate API requests. See API authentication.
  • Binding (secret map) — A rule mapping a stored secret to a specific environment variable on a workload. See manage secrets.
  • Break-glass — A temporary, emergency access request used by platform administrators to inspect project settings. Every request requires a justification and is logged in the project audit log. See break-glass and audit.
  • Build — The automated process that compiles your uploaded source code into a container image. Output details appear in the resource's message field. See deploy an agent.

C​

  • Chunk — A segment of text extracted from a larger document, turned into a vector embedding, and retrieved during search operations. See the RAG chatbot tutorial.
  • CloudEvent — A standardized JSON event format used by the platform to deliver event notifications over HTTP. See HTTP and events.
  • Cold start — The brief delay that occurs when a request arrives for an idle workload, during which the platform initializes a new container instance. See autoscaling and scale to zero.
  • Collection — The internal database collection created behind a VectorDB index. See Indexes and points.
  • Command palette — The quick search modal (Cmd+K or Ctrl+K) in the web console for jumping between resources and projects. See console overview.
  • Concurrency — The number of simultaneous requests a single workload instance processes before the platform scales up additional instances. See autoscaling and scale to zero.
  • Condition — A status check evaluated by the platform to determine overall resource readiness. See deploy a service.
  • Console — The web management portal located at https://console.codyhill.dev. See console overview.
  • Container image — A packaged software bundle containing application code, libraries, and dependencies required for execution. See serverless overview.

D​

  • Data plane — The execution path used to interact with deployed workloads (such as invoking an agent or calling a function endpoint). See invoke.
  • Dead-letter topic — A secondary Pub/Sub topic where messages that fail repeatedly are safely routed for analysis. See topics and subscriptions.
  • Digest — A unique cryptographic hash identifying a specific container image build. See versions and rollback.
  • Dimensions — The fixed number of vector values in each embedding array stored in a VectorDB index. See Indexes and points.
  • Distance metric — The mathematical formula used by VectorDB to calculate similarity between vectors (cosine, dot, or euclid). See search.
  • DNS label — A string format suitable for hostnames, using lowercase letters, numbers, and hyphens. See troubleshooting.

E​

  • Embed key — A public identifier (starting with emb_) used to embed chat widgets on authorized external websites. See embed chat.
  • Embedding — A numerical vector array representing the semantic meaning of text, generated by an AI model. See VectorDB overview.
  • Eviction policy — The rule deciding how MemoryStore manages memory when usage reaches its configured limit. See MemoryStore overview.
  • External URL — The public web address assigned to a published workload, formatted as https://<name>-<project-short>.apps.codyhill.dev. See public endpoints and domains.

F​

  • Framework — The software framework used to write an agent (ADK, LangGraph, or CrewAI). See Agents overview.
  • Function — A single-file serverless handler written in Python, Node.js, Go, or Ruby that scales to zero automatically. See functions overview.

G​

  • Grant — A permission role (member or admin) assigning project access to a user or service account. See projects and access.

H​

  • Harness — The internal web wrapper that loads your agent code and exposes standardized HTTP endpoints. See Agents overview.

I​

  • Idempotent — An operation that produces the exact same outcome whether executed once or multiple times. See publish and consume.
  • Index — A named collection in VectorDB storing vectors that share the same dimension size and distance metric. See VectorDB quickstart.
  • Instance — A running container copy of a deployed workload, or one member of a managed database. See serverless overview.
  • Invitation — A secure link generated by an admin to invite a new user or grant project permissions. See projects and access.
  • Invoke — Sending an HTTP request to execute a deployed agent or function. See invoke.

J​

  • JSON-RPC — The lightweight protocol used by Model Context Protocol (MCP) servers to handle tool requests. See the weather tools tutorial.

M​

  • Management plane — The platform controls used to manage resources, deploy code, set secrets, and inspect logs. See API authentication.
  • MCP client — An application or agent that connects to Model Context Protocol (MCP) servers to execute tools. See connect agents and clients.
  • MCP server — A hosted endpoint that exposes tools to AI models using the open Model Context Protocol standard. See MCP servers overview.
  • Member — A project access role that permits standard operational tasks while restricting administrative deletions and security key reveals. See projects and access.
  • Memorize — The process of extracting facts from a conversation session and saving them into an agent's memory bank. See memory.
  • Memory bank — An agent's searchable vector repository for long-term fact retention across separate conversations. See memory.
  • MemoryStore — Managed in-memory key-value storage that speaks the Redis wire protocol, so any Redis client works unchanged. See MemoryStore overview.

N​

  • Project network space — The private, isolated internal network created for each project to connect services securely.

O​

  • Organization (org) — The top-level account structure owning projects, users, and billing settings. See projects and access.
  • Owner — The user or account responsible for creating a resource or project. See projects and access.

P​

  • Page token — An opaque pagination marker returned by list APIs to fetch the next set of results. See API overview.
  • Payload — The custom JSON metadata attached to a vector entry in VectorDB. See Indexes and points.
  • Persistence — Disk storage configurations that ensure database data survives container restarts. See MemoryStore overview.
  • Platform admin — An administrator managing core platform infrastructure and organization settings. See break-glass and audit.
  • Point — A single record in VectorDB containing an ID, a vector array, and optional metadata payload. See Indexes and points.
  • Poison message — A message formatted in a way that causes handler errors every time it is processed. See topics and subscriptions.
  • Project — A dedicated workspace holding your applications, compute resources, databases, secrets, and team permissions. See projects and access.
  • Publish — Sending a message to a Pub/Sub topic or making a workload publicly accessible on the internet. See publish and consume.
  • Pull — Fetching messages directly from a Pub/Sub subscription over HTTP or gRPC. See publish and consume.
  • Push delivery — Automated delivery where Pub/Sub posts incoming topic messages directly to your application endpoint. See topics and subscriptions.

Q​

  • Quantization — Vector compression techniques that reduce memory requirements while maintaining search accuracy. See Indexes and points.
  • Quota — Resource limits assigned to a project to manage instance counts, storage, CPU, and memory usage. See quotas and audit.

R​

  • RAG (retrieval-augmented generation) — An AI design pattern that fetches relevant document passages from a vector database and provides them as context to an LLM. See the RAG chatbot tutorial.
  • Rate limit — Standard throttling controls that limit request frequency per user or IP address. See troubleshooting.
  • Ready — The boolean field indicating whether a resource is active and ready to accept requests. See API overview.
  • Registry — A repository storing container images for deployment. See deploy a service.
  • Request ID — A unique identifier generated for every API request and returned in error bodies or X-Request-Id headers. See get help.
  • Revision — An immutable, versioned release snapshot of a workload's code and configuration settings. See traffic and revisions.
  • Rotate — Generating a new credential key while invalidating or transitioning away from an old key. See manage secrets.
  • Runtime — The execution environment used for functions (python, nodejs, go, or ruby). See runtimes.

S​

  • Sandbox — An isolated, secure container environment used to execute untrusted code safely. See code sandbox overview.
  • Scale to zero — Automatically shutting down active instances when no incoming requests arrive to optimize compute costs. See autoscaling and scale to zero.
  • Secret — An encrypted configuration value (such as an API token or password) managed securely by Secrets Manager. See Secrets Manager overview.
  • Serverless service — A containerized application deployed as a scale-to-zero HTTPS service. See serverless overview.
  • Service account — A non-human machine account used by automated scripts and CI/CD pipelines. See service accounts and API keys.
  • Session — A tracked conversation thread between a user and an AI agent. See sessions.
  • Subscription — A durable message queue configured on a Pub/Sub topic to track message delivery to consumers. See topics and subscriptions.

T​

  • Topic — A named destination channel in Pub/Sub where publishers send messages. See topics and subscriptions.
  • Trigger — An automated rule that invokes a workload when a scheduled time arrives or an event occurs. See serverless overview.

V​

  • Vector — An array of numbers generated by an embedding model representing data semantics. See VectorDB overview.
  • VectorDB — High-performance vector database service for similarity searches and RAG pipelines. See VectorDB overview.