Glossary

Short definitions of the terms used across this documentation.

Agent / agentic AI

An LLM that doesn’t just answer but can decide to call tools, read retrieved context, and iterate over several steps to reach a result. The AI Agent connector runs one such agent per service task.

Agent loop

The cycle of: send context to the model → model optionally calls a tool → tool result is fed back → repeat → final answer. See Tools.

LLM (Large Language Model)

The text model behind the agent (e.g. an OpenAI/Azure/Ollama model), reached over an OpenAI-compatible API.

OpenAI-compatible API

The request/response shape (/v1/chat/completions, and the Responses API) that many providers implement, letting you switch providers by changing baseUrl/model. See Configuring the Agent.

System prompt / instruction

The standing guidance given to the model. The connector ships a default; your instruction replaces or augments it via instructionMode. See Configuring the Agent.

Tool / function calling

A function the model may invoke (a Java @Tool class or an MCP tool). See Tools.

MCP (Model Context Protocol)

An open protocol for exposing tools to an agent over HTTP; configure servers in mcpServers. See Tools.

ProcessStarterTool

The built-in tool that lets the agent start a CIB seven process by key and read its result. See Tools.

RAG (Retrieval-Augmented Generation)

Grounding answers in your own documents by retrieving relevant chunks and adding them to the prompt. See RAG.

Embedding / embedding model

A numeric vector representation of text; the embedding model turns text into vectors for similarity search. Local default: AllMiniLmL6V2 (384-dim).

Vector / dimension

The embedding’s array of numbers; its length (dimension) must match between ingestion and query. See RAG.

pgvector

The PostgreSQL extension that stores embeddings and does similarity search; the RAG store. See RAG.

Chunk / chunk overlap

Documents are split into overlapping text chunks before embedding (chunkSize/chunkOverlap). See RAG.

Chat memory / memoryId

Retained conversation history that lets a process hold a multi-turn dialogue across steps, keyed by memoryId. See Chat Memory.

Reasoning model / reasoning effort, summary

Models that “think” before answering; reasoningEffort hints the budget, reasoningSummary exposes a summary (via the Responses API). See Configuring the Agent.

Responses API

An OpenAI API variant used when reasoningSummary is set.

Connector / connect plugin / element template

A connector is the engine integration (cibseven-ai-agent); the connect plugin is how it’s packaged; the element template is the Modeler form that configures it. See Overview and Installation.

aiMeta

The machine-readable marker tagging output as AI-generated (provenance). See Audit Trail.

Chat-log (audit) variable

The per-activity process variable holding the audit timeline of an agent invocation. See Audit Trail.

Redaction

Replacing stored message content with a hash+length in the audit log; protects the stored copy, not what was sent to the provider. See Audit Trail and Security & Data Handling.

Token

The unit LLM providers measure (and bill) input/output in; more context ⇒ more tokens ⇒ more cost/latency. See Overview.

Prompt injection

Adversarial instructions hidden in content the agent reads (a document, a tool result), aiming to steer its behaviour. See Security & Data Handling.

Hallucination

A confident but unsupported model claim; mitigated by RAG grounding and by validating output. See RAG and Getting Started.