Demos¶
Tryll ships two demo projects. Each one is a game project with ready-made agents that you can talk to in the editor and open in the graph editor.
| Project | Engine | Contents |
|---|---|---|
TryllUnityDemo |
Unity 6000.3 or later | A gallery of 11 agents, one workflow pattern each |
TryllUnrealDemo |
Unreal Engine 5.7 | The same gallery in Unreal, 10 agents |
The release package for each engine holds the demo projects next to the plugin
folder. Each demo loads the plugin from that folder, so a demo project must stay
next to it. Each project's README.md says how to point it at the plugin if you
move it.
In the tables below, Nodes lists the nodes in the order a turn reaches them. A slash separates the two routes after a branch.
Unity demo project¶
Open and run it.
- Add the
TryllUnityDemofolder to Unity Hub and open it with Unity 6000.3 or later. - Set up the server and models as in First Inference in Unity.
- Open
Assets/DemoScene.unity. Each demo agent is a GameObject under Agents in the Hierarchy. - Open Window → Tryll → Chat and assign the agent's GameObject to Agent Component. See Test an Agent in the Editor.
Each agent's GameObject has a Scenario Note component. In the Inspector it
shows the pattern the agent demonstrates and the lines to type to test it. The
workflow assets are in Assets/Workflows/. Every agent except Emotional Voice
Character has a dialog script next to its workflow (<Name>_Dialog.asset) for the
Dialog Lab (Window → Tryll → Dialog Lab).
The microphone in the voice demos works only in Play Mode. Aquarium Expert needs the embedding model BGE Small EN v1.5 downloaded first in Window → Tryll → Model Manager, because Tryll does not download embedding models for file-backed storages automatically.
| Demo | Character | What it shows | Nodes | Video |
|---|---|---|---|---|
| Safe Receptionist | Mr Alvarez, front-desk receptionist at the Hotel Meridian | A regex check sends known jailbreak and insult phrasings to prewritten replies, and every other line to the model. | Regex Guardrail → Canned Response / Generate | Watch |
| Aquarium Expert | An aquarium fish-keeping expert | The model answers from facts retrieved from a knowledge base, and a prewritten refusal answers questions the knowledge base does not cover. The Advanced variable switches to a hidden query-rewriting step and hybrid search. The scene holds two agents: AquariumExpert and AquariumExpert-Advanced. |
Branch → (Generate → Retrieve) / Retrieve → Transform → Generate / Canned Response | Simple, advanced |
| Quest Giver | Mira, quest giver of Bramblehollow village | Agent variables (player name, reputation, completed quests, village mood) are written into the prompt each turn. Two agents, QuestGiver-Newcomer and QuestGiver-Hero, share one workflow and differ only in their variable overrides. |
Transform → Generate | Watch |
| Station Merchant | A merchant on orbital station Delta-IV | The model calls a get_price tool, game code returns the price, and the model quotes it. |
Tool Call → Generate | |
| Voice Character | Sasha Glotov, CEO of Tryll Engine | The player speaks into the microphone and the character answers aloud, sentence by sentence, with a Supertonic voice. | Generate and Speak | Watch |
| Emotional Voice Character | Kira, a K-pop singer interviewed by a magazine | A cloned voice with seven delivery modes, one per emotion. The scene sets the flirty mode. |
Generate and Speak | |
| Intent Guard | Aldric, a grumpy gate guard | An embedding classifier matches the player's line to an intent (insult, pleasing, joking) and adds a tone instruction for that intent to the reply prompt. | Classify Intent → Intent to Instruction → Transform → Generate | |
| Intent Guard (LLM) | Aldric, a grumpy gate guard | The same guard as Intent Guard. A small language model picks the intent from written descriptions, so the demo needs no example phrases and no embedding model. | Classify Intent (LLM) → Intent to Instruction → Transform → Generate | |
| Command Parser | None: the agent turns a request into a game command | A grammar limits the reply to one of 16 two-word commands, such as pull lever, so game code can split it into action and object without error handling. |
Generate | |
| Tavern Bartender | Tobin Tallow, barkeep in the mining town of Orvyn | Every 7 turns, a Summarize node folds older turns into a rolling summary that keeps what the player said. | Generate → Summarize | |
| Immersion Guard | A bread seller in the market square of Orvyn | A regex check, then one classifier call that scores two things: whether the line is aimed at the software, and whether it belongs in a medieval world. A line caught by either check gets a prewritten deflection. | Regex Guardrail → Classify Attributes (LLM) → Generate / Canned Response |
Unreal demo project¶
Open and run it.
- Open
TryllUnrealDemo.uprojectwith Unreal Engine 5.7. It is a C++ project, so the editor offers to build its modules on first open. The build needs Visual Studio with the C++ game development workload. - Set up the server and models as in First Inference in Unreal.
- The editor opens the map
M_Gallery. It holds one actor per demo agent, named after the demo. - Open Window → Tryll → Tryll Chat and pick the demo's actor in Actor. See Test an Agent in the Editor.
The workflows are in Content/Workflows/ and the dialog scripts for the
Dialog Lab (Window → Tryll → Tryll
Dialog Lab) are in Content/DialogScripts/.
The gallery has the same agents as the Unity demo project, with the same characters and nodes, except Emotional Voice Character. Quest Giver and Aquarium Expert have one actor each.