Concepts¶
These pages explain how Tryll thinks — the mental models behind the server, the client libraries, and the workflows you build on top of them. Read one when you want to understand why something works the way it does; reach for a how-to guide or the reference when you want to do a specific thing or look up a specific field.
Start with these¶
- Architecture at a Glance — the 10-minute tour. Server → session → agent → graph → node → model. Every later page fits inside this frame.
- Workflows, Graphs, and Nodes — the building block that turns a model into an application.
- Agents and Sessions — who owns what, and how a conversation comes into and goes out of existence.
- Lifetime and Ownership — how string storages, embedded string storages, and models stay alive across agents, turns, and sessions.
Dive into the pieces¶
- Models and Inference Engines — language vs. embedding models, variants per backend, what the KV cache does.
- TTS and Voice Output — how streaming
speech synthesis works:
GenerateAndSpeakvs.Speak, PCM audio frames, voices, and the Supertonic / Pocket TTS families. - STT and Voice Input — how spoken audio becomes text: the utterance lifecycle, push-to-talk vs. hands-free, offline vs. streaming recognisers, and auto-sending into an agent.
- Retrieval-Augmented Generation (RAG) — how Tryll grounds answers in your data: chunking, embeddings, cosine distance, and where retrieved text lands in the prompt.
- Tool Calling — making small local models invoke tools: chat-template-driven prompting, how results flow back into the dialogue, and the limits that matter on-device.
- Agent Variables — a typed, per-agent
key→value store the game writes over the wire and the server reads
from Mustache templates (
{{var.<name>}}), the Retrieve filter, and (opt-in)__NAME__markers in Generate / GenerateAndSpeak output. - Slots and Inter-node Value Passing
— the per-turn blackboard nodes read and write:
inputselection,{{slot.<name>}}templates, andsend/history_roledelivery. - Constrained Output (GBNF) — forcing a model to emit exactly a fixed shape (a command, a choice, a tiny JSON object) with a grammar: what it guarantees (syntax) and what it never does (meaning).
- Projection and Token Budgets — how a growing dialog is turned into a bounded prompt, turn after turn, without losing the session.
How these relate to the rest of the docs¶
flowchart LR
C[Concepts<br>why & mental model] --> R[Reference<br>fields & semantics]
C --> H[How-to<br>task recipes]
H --> R
T[Tutorials<br>first inference] --> H
T --> C
Every concept page links down to the specific reference pages that make its ideas concrete, and to the how-to guides that put them to use.