This page is machine-translated from French. Read the French original.

Glossary

Agent

An assistant authorized to act : to call upon tools, to link steps together, to produce a concrete result. Triggered by @agent in a conversation. → Agents Verification Status

Vector base

The database stores the vectors of document fragments and retrieves those closest to a question. → Vector base and embeddings

Skills library

A set of job instruction sheets, loaded on demand by an agent. This is distinct from skills, which involve executing code. → Skills library

Skill

A tool that an agent can call upon: web search, file reading, SQL query, calendar… Activated by an administrator, then by workspace. → SKILLS

Context (window of)

The amount of text a model can process at one time. This includes the prompt system, pinned documents, retrieved passages, history, and query. When it is full, the oldest items are discarded.

Cutting

The splitting of a document into fragments of homogeneous size, with an overlap, to allow for searching.

Embedding

Transforming a text into a vector — a list of numbers representing its meaning. Two texts with similar meanings produce similar vectors.

Workspace

WivenLLM's organizational unit: a collection of documents, settings, and conversations. Two spaces share nothing. → Workspaces

Pinned (document)

An injected document in full within the context of each message, without needing to search. Suitable only for short and consistently relevant documents.

Thread

An independent conversation within a workspace. Threads share documents and workspace settings, but not their history.

Agent flow

A sequence of steps defined in advance and always executed in the same order. Unlike an agent, who decides their own path. → Agent flow

Fragment (chunk)

A piece of document resulting from the segmentation. This is the unit that the search finds and which is provided to the model.

Hallucination

A statement confidently produced by a model, but false or fabricated. Query mode and source citations are used to contain and detect it.

LLM (Large Language Model)

The engine that writes the answers. WivenLLM is not one: it orchestrates the one you choose.

MCP (Model Context Protocol)

An open standard describing how an AI application discovers and uses external tools. → MCP Servers

Memory

A short, memorable fact about you, then used in conversations. → Memoirs

Chat Mode

The assistant uses your documents And his general knowledge.

Query Mode

The assistant replies uniquely from your documents, and explicitly rejects them when they do not contain the answer.

Multi-user mode

The mode that enables named accounts, roles, shared spaces, web widget and API.

Prompt system

The permanent guidelines given to the model before each message: role, tone, format, prohibitions. The application's most effective lever. → Conversations

RAG (Research-Augmented Generation)

The central mechanism: find the relevant passages in your documents, provide them to the model, then generate the answer from them.

Model router

A mechanism that automatically directs each conversation to the most appropriate template, according to specific rules. → Model router

Second Brain

Your personal knowledge base, which spans all workspaces. → Second Brain

Similarity threshold

The minimum degree of proximity a fragment must reach to be retained by the search. Too high, it excludes useful information; too low, it lets in noise.

Sources

The document excerpts displayed below an answer indicate what it is based on. They allow for verification.

Planned task

An agent's instruction executed automatically according to a schedule. → Planned tasks

Temperature

The degree of unpredictability of the responses. Low for factual answers, high for creative answers.

Web widget

A chat bubble to insert into a website, linked to a workspace. → Web widget