Manage Models in the Editor¶
Both the Unity and Unreal plugins include an editor-only Model Manager window for working with the server's model catalog without writing any wire code: browse every model the server knows about, download and load them, free disk space, and mark which models ship in your packaged build.
It is a graphical front end over the model-management
wire protocol (ListModels, DownloadModel,
LoadModel / UnloadModel, DeleteModel). The catalog itself is defined
server-side in models.json.
Prerequisites
- The Tryll plugin installed (see First Inference in Unity / Unreal).
- A
TryllRuntimeSettingsasset (Unity) or configured plugin settings (Unreal). The window launches the bundled server automatically, the same way the Chat Window does.
Step 1 — open the window¶
Window → Tryll → Model Manager
The window opens with the title Tryll Models and connects to the server automatically.
Tools → Tryll → Tryll Model Manager
Opens as a dockable tab (Tryll Model Manager) and connects on open.
The layout is the same in both engines: a filters column on the left, the model list in the middle, and a properties + actions panel on the right.
Step 2 — find a model¶
Use the left column to narrow the list:
| Filter | Values |
|---|---|
| Search | Match by model name or HuggingFace repo. |
| Model type | Language, Embedding, STT, TTS, VAD. |
| Engine | llama-cpp, sherpa-onnx. |
| Status | Absent, Downloaded, Local, Loaded, Downloading. |
| Registration | Production, Experimental, Not registered. |
The left column also totals how many models are registered for the build and the on-disk / shipped size, so you can see your build's model footprint at a glance. Select a row to populate the properties panel on the right.
Step 3 — download, load, unload, delete¶
Select a model, then use the action buttons in the properties panel:
| Action | Effect | Wire call |
|---|---|---|
| Download | Fetch the model's files from HuggingFace; a progress bar tracks bytes. | DownloadModel |
| Load | Pin a language / embedding / STT / TTS model into server memory so the first turn has no load latency. VAD cannot be pinned (catalog/download only). | LoadModel |
| Unload | Unpin a loaded managed model (freed once no agent still uses it). | UnloadModel |
| Delete files | Remove the downloaded files from disk. | DeleteModel |
A toolbar Download all action fetches every still-absent model in the current
filtered list. See
Model Management → Lifecycle for the
load/unload rules (you usually do not need to Load manually — an agent loads
its model on demand at CreateAgent).
Step 4 — register models for your build¶
Each model can be tagged with a build registration tier, in the properties panel:
| Tier | Meaning |
|---|---|
| Production | Shipped in your packaged build. |
| Experimental | Available in the editor; stripped from packaged builds. |
| Not registered | Not part of the build at all. |
Registrations are stored in your project, not on the server:
Persisted to a TryllModelManifest asset
(Assets/Tryll/TryllModelManifest.asset).
Persisted to DefaultGame.ini via the Tryll Model Manifest developer
settings.
Step 5 — pick registered models on nodes¶
Model-name fields elsewhere in the editor read this manifest. A Generate node's
model, GenerateAndSpeak's TTS model, and the Voice Input component's
STT model all show a dropdown filtered to registered models of the right kind,
with a tier badge next to each. A name you set that is not registered still
displays — it is never silently dropped — but it will not ship unless you register
it Production.
This is the editor side of the (tryll_model) schema attribute; see
Model Management → Build registration.
Common problems¶
- The list is empty / "Connecting…" never resolves
- The server did not start or did not bind port 9100. Check the Unity Console or
Unreal Output Log for
[Tryll Server]lines, as in the Chat Window troubleshooting. - A model shows
Absentand Download is disabled - The catalog entry has no
huggingface_repo, so it cannot be fetched over the wire — it must be supplied on disk. See Use Your Own Local Model. - A registered model is missing from a build
- Only Production models ship. Experimental and unregistered models are excluded from packaged builds by design.
Related¶
- Reference: Model Management
- How-to: Use Your Own Local Model
- How-to: Pin and Unpin Models
- How-to: Verify Required Models Before Play — the editor guard that catches a registered-but-undownloaded model before Play/PIE or Tryll Chat.
- Concept: Models and inference engines