Glossary

What the terms in this wiki mean, and the page that explains each one.

A term links to its entry here the first time it appears in the wiki.

Engine#

The program that loads a model and writes its replies. The Draggy engine runs GGUF models on your computer. See Providers.

colibri#

The engine Draggy uses for large mixture-of-experts models that do not fit in memory. It reads parts of the model from the drive as it goes, so it is slow. See Providers.

Provider#

A source of models in Draggy: the built-in engine, a server on this computer, a cloud provider with an API key, or a ChatGPT or Claude plan. See Providers.

GGUF#

The file format of the models the Draggy engine runs. A model is one .gguf file, or several for very large models. See Choosing a model.

Quantization#

A way of storing a model's numbers in fewer bits, which makes the file smaller. Settings → Models lists each variant of a model with its file size. See Choosing a model.

Context window#

The amount of text, counted in tokens, that a model can hold at once. The context wheel in the composer shows how full it is. See Chat and Code.

Compaction#

Folding the older part of a long conversation into a summary, so the conversation can carry on. It happens at the compaction limit: 60% of the context window by default, and at most 64,000 tokens for paid cloud models unless you set a limit. See Chat and Code.

MCP server#

A program that gives Draggy more tools, through the Model Context Protocol. Settings → Extensions lists them. See Extensions and skills.

Skill#

A folder with a SKILL.md file that teaches the model one kind of job. The model reads a skill only when a request fits it, and you can start one by typing / and its name. See Extensions and skills.

Embedding model#

A model that turns text into vectors, which the Library uses to search documents by meaning. It cannot hold a conversation. See Choosing a model.

Talk#

Draggy's spoken conversation mode, in the Chat sidebar. It listens, works out when you have finished, answers out loud, and lets you cut in. See Voice mode.

YOLO (permission mode)#

The permission mode that edits files and runs commands without asking. Keep it for projects you can restore. Web search has an Automatic setting too, and it does not change permissions. See Permissions.