Skip to main content
This is the shared vocabulary used across the docs. Each concept links to the page that covers it in depth.

​
Agents and topology

Agent: a conversational program made of prompt nodes wired together, with one node as the entrypoint. The same agent serves both chat and voice. Node: the core primitive, a single-prompt step with a stable key, a label and description, a model binding, a system prompt, and a list of tools. Topology: the graph of nodes that makes up an agent. A conversation traverses one path through it; different callers take different paths. Entrypoint: the node every conversation starts on. Its published version pins the whole reachable graph. Handoff (agent_swap): how control moves between nodes. A handoff is a callable tool on a node that targets another node as (target_key, target_version). If it is not present as a tool, the agent cannot take that path. See Agents.

​
Tools

Tool: anything the model can call on a node. There are four shapes:
  • agent_swap: move control to another node (a handoff).
  • end_call: end the conversation.
  • custom: call code your organization deployed.
  • http_req: make a declarative external HTTP request.
Custom tool: your own code, deployed to Gradient and run by the engine as a tool. It has a slug, name, description, JSON input schema, runtime, source, deploy status, an endpoint URL, and an immutable deploy history. Authored in the Tools tab. See Custom tools. See Tools.

​
Knowledge and skills

Knowledge base: a collection of documents, chunked and embedded, that a node can search to ground its answers in your content (retrieval). Skill: a reusable block of instructions you write once and attach to any node, instead of copying the same guidance into many prompts. See Knowledge and skills.

​
Versions

Draft: the single mutable builder graph. Every edit lands here. It is never a numbered version and keeps that identity forever. Published version: an immutable snapshot minted by publishing, which copies the reachable draft into the next positive integer (1, 2, and so on). Live traffic resolves to a published version; restoring copies a published revision back over the draft. Each node versions independently, but the entrypoint version pins the whole reachable graph. See Versions.

​
Data

Dataset: a Postgres database holding your agent’s world: patients, orders, accounts, whatever the tools read and write. Each organization gets a managed Postgres project, and each dataset is a database inside it, browsable as tables with a SQL console. Snapshot: a pinned point in a dataset’s history. Forks taken from it start from a known state. Fork: a private, writable copy of a dataset created for a single conversation. Tool calls read and write the fork, so multi-step scenarios keep state, and the real dataset is never touched. Forks power the redteam. See Datasets.

​
Evaluation

Rubric: a topology’s definition of correct, a set of scenarios, each carrying rules. The same rubric grades both redteam and production conversations. Scenario: one concrete situation the agent is meant to handle, bound to specific dataset rows so it is reproducible. Rule: an LLM-judged assertion on a scenario, written as a sentence. A judge model reads the transcript and the tool calls and returns pass, fail, or skipped. Scoring excludes skips: pass divided by (pass plus fail). Trace: the full record of one conversation, every turn, every handoff, and every tool call with its arguments and result. Chat and voice both produce traces, and rubrics grade them. See Results. See Rubrics.