Glossary

Short definitions for every concept used across the docs. The detail lives in linked pages — this is the lookup index.

Authoring concepts

Agent. A named identity composed of Skills and Workflows, with a system prompt and memory, grounded in Sources. The agent routes work to its skills; it is not itself a tool set. Authored in workspace/<org>/agents/<slug>.yaml. See Agents.

Skill. One of an agent's roles — a responsibility that bundles its logic (a typed operation), its acceptance criteria (evals), and the tools it may call. A skill is not a tool; it decides which tools to use and is held to an acceptance bar. Authored in workspace/<org>/operations/<slug>/. See Skills. SDK note: the public TypeScript type is Operation (@vocion/sdk's defineOperation) — kept stable as a plugin contract — but the user-facing name everywhere in the dashboard, marketing, and docs is "Skill." The dashboard route is /dashboard/skills as of vocion-v0.5.1. The brief v0.5.0 experiment with "Operations" as the UI label was reverted.

Tool. An atomic capability a skill calls — search_knowledge, lookup_objects, run_operation, request_human_review, and the learnings/runs tools. The lowest layer of the intelligence stack (Agent → Skill → Tool). Recorded in the run trace. Custom tools ship via plugins. See Tools.

Workflow. A sequence of steps — Skill, action, or approve gate — with persistent state. HITL pauses on approve steps. Authored in workspace/<org>/workflows/<slug>.yaml. See Workflows.

Object. A first-class business entity (Account, Opportunity, Deck, Transcript) the tenant cares about. Typed via workspace/<org>/objects/<type>.yaml. See Objects.

Source. A connected system feeding documents into the retrieval layer (Google Drive, GitHub, Slack, web crawl, file upload). Implemented as a typed plugin against @vocion/sdk. See Sources. Note: the dashboard label is "Sources" everywhere as of vocion-v0.5.0 — the prior "Connectors" UI synonym was retired.

Playbook. Markdown + YAML procedural guide mounted into agent context at /playbooks/<slug>/. Authored in workspace/<org>/playbooks/. See Playbooks.

Learning. Whitelisted rule-step entry that agents can commit during runs (via add_learning). Authored in workspace/<org>/learnings/ and surfaced to the agent via virtual file mounts. See Learnings.

Eval. A scored dataset of input → expected-shape pairs that exercises an Agent or Skill in CI. Stored in workspace/<org>/evals/<slug>.yaml. See Evals concept + Evals guide.

Runtime concepts

Run. A single execution record. skill_run for Skill executions, workflow_run for workflows. Both stamp the active workspace_sha for traceability.

workspace_sha. Git-style hash of the active workspace/<org>/ snapshot at run time. Stamped on every run so any output can be replayed against the exact prompts that produced it.

Retrieval. Looking up relevant chunks from indexed knowledge sources before passing them to an LLM. Vocion uses Postgres + pgvector (vector) and tsvector (keyword), fused via RRF. See Retrieval config.

Embedding. A numerical vector representation of a chunk of text, produced by an embedding model (OpenAI text-embedding-3-small, 1536-d by default). Stored on knowledge_chunk.embedding.

Chunk. A token-aware slice (~512 tokens by default, 64-token overlap) of a source document. The unit Vocion embeds and retrieves.

RAG. Retrieval-Augmented Generation — the pattern of retrieving relevant chunks first, then having the LLM answer with that context. Vocion's default agent loop runs RAG via the searchDocs / searchKnowledge tools.

Hybrid retrieval. Combining vector (semantic) and keyword (lexical) results via RRF — Reciprocal Rank Fusion. Better recall than either alone on typical docs corpora.

Plugin. A typed TypeScript module that registers a Skill or Source via @vocion/sdk. Loaded at startup; takes precedence over a prompt Skill with the same slug. See Writing a plugin.

Connector. Historical synonym for Source. Retired from the dashboard UI in vocion-v0.5.0 — the framework code, marketing copy, and the dashboard nav now all say "Source." Use Source.

MCP. Model Context Protocol — Anthropic-defined RPC standard for LLM tools. The Vocion MCP server exposes every Agent, Skill, Workflow, and Object as a callable tool. See MCP reference.

Governance + auth

RBAC. Role-based access control. Roles: org:admin, org:member. Defined per organization. See authentication.

DBAC. Data-based access control. Per-row ownership and scope on every domain table (org_id everywhere; user-scoped fields where relevant). Cross-tenant queries forbidden at the service layer.

Org / Tenant. A Vocion organization. Vocion is multi-tenant by org; every domain row is org_id-scoped.

Observability

Trace. A Langfuse record of one logical LLM operation. Every Vocion run produces a trace tagged with feature:<name>, org:<orgId>, slug:<slug>, userId. See Observability.

Feature. Closed enum (agent.chat, operation.run, eval.judge, workflow.step, retrieval.search, …) stamped on every trace. Lets you slice cost and volume by feature in the Langfuse UI.

Budget. Per-agent token + USD cap stored in agent_budget. Pre-flight check refuses runs over the hard cap. See Evals + budgets.