Video Walkthroughs¶

Screen recordings that build and test the demo agents from the Unity demo project (TryllUnityDemo), one node at a time. Each video opens the workflow in the graph editor, explains the parameters that matter, talks to the agent in Tryll Chat, and then tests it in the Dialog Lab. The nodes and parameters are the same in Unreal; only the editor windows look different.

The whole series is also a YouTube playlist: Tryll documentation videos.

Video	Length	What it shows
Safe Receptionist	9:21	Guardrails, canned responses, and testing with the Dialog Lab
Voice Character	4:43	Voice output with Generate and Speak, voice input outside the graph
Quest Giver	6:36	Agent variables, the Transform node, and slots
Aquarium Expert	9:20	The Branch node and simple RAG with a Retrieve node
Aquarium Expert, advanced path	13:27	Query rewriting, HyDE, and hybrid retrieval

Watch them in order: each video builds on the ones before it.

Safe Receptionist¶

The simplest useful agent: a hotel receptionist that answers guests and refuses what it should not answer. Three nodes and one branch.

The graph. A Regex Guardrail node checks the player's message first. On triggered the turn goes to a Canned Response node; on not_triggered it goes to a Generate node.
The guardrail. A short inline list of regular expressions, and when to move the list into a string storage file instead.
The refusal. Selection Strategy, why a canned line is sent whole rather than streamed, and what History Role does on the next turn.
The model. Why Generate nodes leave Model Name empty and use the workflow's Default Model Name, and why a prompt-level rule is weaker than the regex node.
Testing. Why one chat window is not enough. A dialog script that checks both routes, including a false positive check on a harmless question. Five repeats of one variant to see the spread, then three models side by side.

Voice Character¶

An agent you talk to and that talks back. The whole workflow is one node.

Voice out. One Generate and Speak node. It speaks each sentence as soon as it is finished, while the model writes the next one. Min Sentence Chars trades time to first audio against more, smaller synthesis calls.
The voice. The TTS model (Supertonic), picking a voice by Speaker Id, Speed, and which voice fields a fixed-voice model ignores.
A prompt written to be spoken. Short plain sentences, no lists, no markdown, no stage directions.
The scene. The agent component, the speaker (attached for you if you leave it empty), and the voice input component.
Voice in. Speech recognition is not a node: the voice input component owns the microphone and sends the transcript to its Target Agent as if the player had typed it. Voice-activity detection, a hotwords file, and hands-free versus push-to-talk (the turn starts when you let go of the button).
What the Dialog Lab covers. The Lab tests the text half of a speaking agent. Microphone, silence detection and playback still need a human listening.

Quest Giver¶

One graph, one model, one system prompt, and a quest giver who treats a stranger very differently from the hero who saved the village. None of that difference lives in the graph: it comes from the game.

Variables. Declaring typed agent variables on the workflow (a string, an integer, a string set), and why the declared value is only the starting value.
One asset, two NPCs. Variable Overrides on the agent component, like a material and a material instance in Unity.
Slots. The one-turn blackboard, and how a slot differs from a variable. See Slots and Inter-node Value Passing.
The Transform node. A Transform node renders a Mustache template with the game state and the player's message into a slot. No model runs and nothing reaches the player.
Input. The Generate node's Input reads that slot instead of the raw message, so the model always sees the question wrapped in the state of the world.
Testing. Two Dialog Lab variants that differ only by agent, the two columns side by side, and the rendered prompt for each.

Aquarium Expert¶

An expert that answers from a knowledge base instead of from the model's own training: retrieval-augmented generation, the simple version. This video takes the simple path through the workflow; the next one takes the advanced path.

The Branch node. A Branch node reads one boolean variable (Advanced) with the is true test and picks then or else. It runs no model and sends nothing.
How vector search works. Text becomes a vector, similar meanings land close together, and the question is matched by meaning, not by keyword.
The knowledge base. The config file, the records file, the embedding model, and what metadata is for. See Embedded String Storage.
The Retrieve node. The Retrieve node's Input, Retrieval Mode, Top K and Threshold. The threshold is a maximum distance, so a lower number is stricter. found goes on to the answer; not_found goes to a canned response read from a file.
Folding knowledge into the prompt. A Transform node puts the retrieved chunks first and the question last.
Reading the results. The route and the per-node chips in the Turn Inspector, the hit table with distances, and the prompt the model was sent.
Where it fails. An off-topic question correctly misses; a loose threshold lets useless chunks through; a question about a fish that is not in the knowledge base produces a fluent answer from the model's own training. You can prove it is a hallucination because both halves are on screen.

Aquarium Expert, advanced path¶

The same workflow with Advanced turned on: two extra steps that fix the two real flaws of plain vector search.

Flaw one. The player's message is not written for searching: typos move the vector, and "they" or "that one" carry no meaning into an embedding.
Flaw two. Your game's own words (items, places, invented species) are noise to an embedding model that has never seen them. Those need keyword matching.
The rewrite node. A hidden Generate node with Send and History Role both set to None: the player never sees it and later turns never replay it, but it still reads the conversation, which is how it resolves "they".
Query rewriting versus HyDE. Rewrite the question into a search query, or into a hypothetical answer written like the stored chunks. The passage may be wrong; it only has to land next to the right chunk. A Keywords line feeds the keyword search.
The settings that matter. A small Context Size, Temperature 0, and a measured Max Tokens.
Hybrid retrieval. Dense and BM25 searches, why their scores cannot be averaged, and how Reciprocal Rank Fusion combines ranks instead. In hybrid mode the threshold filters only the vector search.
Reading a hybrid hit table, and a payoff turn: a misspelled question that never names the fish, simple path against advanced path.