Solaris

Solaris is a household AI assistant that ServiceBay deploys as one click. Its core is the Solaris Engine — a native agent loop inside solaris-chat that talks directly to a local Ollama, controls the home via Home Assistant, and fronts the Voice PE speaker through HA's Assist pipeline. (The earlier Hermes-gateway architecture was fully replaced in v0.10 — see solaris-architecture.md for the full picture and flows.)

flowchart LR
    PE["🔊 Voice PE"] -- ESPHome --> HA["HA Assist pipeline<br/>whisper GPU · Martin TTS GPU"]
    Browser["💻 Browser"] -- SSO --> Chat
    HA -- "conversation.solaris" --> Chat["Solaris Engine<br/>(solaris-chat)"]
    Chat -- "per-turn model+think" --> Ollama["ollama (GPU)<br/>e2b · 12b"]
    Chat -- "tools · registry · announce" --> HA
    Chat --- DB[("solaris.db")]
    Chat --- Notes[("notes vault")]
    Chat -- "admin only" --> SB["ServiceBay MCP"]

Loading

A spoken command answers in ≈1.3 s after speech end (whisper GPU 0.38 s + engine ≤1 s). The household prompt measured ~7.8k tokens on this box on 2026-08-03 — one household turn against the live HA registry, 34 tools and 51 injected entities. Its composition is logged per prompt assembly as engine.prompt.composition (tools/soul/registry/scaffold), so the number can be re-measured rather than guessed.

Features

Detail docs live in docs/features/; the architecture record is solaris-architecture.md and the design rationale docs/solaris-concept.md. Where the system is headed — zones, ADRs and the V1 backlog — is docs/solaris-zielarchitektur.md.

Assistant

• Voice + Chat, German, GPU-fast. Talk to the Voice PE (wake word "Solaris", on-device) or chat in the browser; a spoken command answers in ≈1.3 s. Two chat modes: Zuhause (fast gemma4:e2b) and Solaris Gründlich (thorough gemma4:12b). → chat-and-voice.md
• Home control via Home Assistant. Lights, covers, media, sensors — with a confirmation gate on locks and garage/gate covers. Chat offers 2–4 quick-reply chips; voice re-opens the mic when the answer ends in a question.
• Music & radio. „Spiele Musik von <Künstler>" plays from Jellyfin; „Spiele Radio" resolves a station and remembers your favorite.
• Timers & reminders. „Stelle einen Timer auf 10 Minuten" — the engine's scheduler rings the speaker back when it fires.

Household surfaces

• Favorites start page (/p/start). Pin devices/actions by voice („packe das Bürolicht auf meine Startseite"), by ☆ tap, or via the room-grouped card picker. Häufig genutzt surfaces what you use most; PWA install + mobile bottom nav (🏠 Zuhause · 💬 Chats · ⭐ Favoriten). → favorites-start-page.md
• Energy page (/p/energy). Live PV / Haus / Netz / Akku flow with correct directions, lifetime kWh totals, per-circuit power, and 24h/7d trend charts. → energy.md
• Notizen tab (/p/notes). Browse / search / read / edit the vault, see stats, and curate the inbox of loose facts. → notes-tab.md
• Concept pages (/c/<id>) render any knowledge entity; #tag / @person mentions in chat auto-link and drive grouping.

Knowledge

• One household knowledge graph. Ask „Wen habe ich letzte Woche gesehen?" — a nightly Stenograph captures facts from your conversations, a Bibliothekar consolidates them, and everything is stored as OKF concept files (people/events/places/…) with a rebuildable solaris.db projection. → knowledge-system.md
• Unified semantic search. okf_vectors + numpy cosine top-k folded into one notes_search tool that blends fuzzy, entity/alias, date-range event, and semantic hits — all default-deny scoped.

Ingest

• Pulls in your on-box data, read-only and idempotent: Obsidian notes, Immich photos, CalDAV/CardDAV calendar & contacts, Jellyfin music, per-person IMAP email, and WhatsApp/Signal/SMS export drops. → ingest.md

What's in this repo

• Solaris Engine + chat surface (solaris-chat/) — one process owning the agent loop (direct Ollama /api/chat, per-turn model + reasoning), the session store (solaris.db), native LLM tracing, the timer scheduler (speaker delivery via assist_satellite.announce), the night crons, the chat UI, and the Ollama-compatible facade HA's conversation agent calls. Built into ghcr.io/mdopp/solaris-chat:latest.
• Skill packs (templates/solaris/skills/) — markdown procedure packs the engine folds into its prompts: household/ (incl. the cron-job bodies daily-chronicle, problem-summarizer) and admin-soul/ (the operator persona: admin-diagnose, admin-logs, admin-act + its SOUL.md).
• ServiceBay templates (templates/{ollama,solaris}/) — two services: ollama (the local LLM engine) and solaris — one Pod with the chat (engine) and gatekeeper containers. post-deploy.py seeds the soul, adopts the HA token, wires the voice pipeline (wyoming whisper/piper, the Solaris conversation agent, the Assist pipeline on the Voice PE) and mints the servicebay_admin MCP token.
• Solaris stack (stacks/solarisbay/stack.yml) — bundles the two templates so a ServiceBay operator can install with one click.
• Voice gatekeeper image source (voice-gatekeeper/) — Python Wyoming-protocol bridge for wyoming-satellite hardware (the Voice PE itself rides HA's Assist pipeline); turns run against the engine's facade. Built into ghcr.io/mdopp/solaris-gatekeeper:latest.
• Database image source (database/) — Alembic schema-init container that runs alembic upgrade head against solaris.db on every pod start. Built into ghcr.io/mdopp/solaris-schema-init:latest.
• Wakeword trainer image source (wakeword-trainer/) — the GPU companion that claims queued wakeword_training_runs and trains the on-device microWakeWord "Solaris" model. The engine can only enqueue; training needs TensorFlow and the GPU. Built into ghcr.io/mdopp/solaris-wakeword-trainer:latest.

Install

1. ServiceBay → Settings → Registries → Add mdopp/solarisbay (https://github.com/mdopp/solarisbay.git).
2. After save, the ollama + solaris templates and the solarisbay stack appear in the wizard.
3. Install the stack. The solaris template's post-deploy.py does the rest (soul, HA token adoption, jellyfin integration, voice pipeline, admin MCP token).

Repository layout

solarisbay/
├── README.md                       # this file
├── solaris-architecture.md         # the architecture record
├── templates/                       # ServiceBay templates
│   ├── ollama/                       # the local LLM engine — its own service
│   └── solaris/                      # the assistant service
│       ├── template.yml             # one Pod: chat (engine) + gatekeeper
│       ├── post-deploy.py           # soul + HA wiring + admin MCP token
│       ├── variables.json
│       └── skills/
│           ├── household/           # household skill pack (engine prompts)
│           └── admin-soul/          # operator skill pack + SOUL.md
├── solaris-chat/                   # Docker image source (the Solaris Engine)
├── voice-gatekeeper/               # Docker image source (Wyoming bridge)
├── database/                       # Docker image source (alembic)
├── wakeword-trainer/               # Docker image source (microWakeWord GPU)
├── stacks/
│   └── solarisbay/
│       └── stack.yml               # templates: [ollama, solaris]
└── .github/workflows/
    └── build-images.yml            # publishes the GHCR images

Image build

.github/workflows/build-images.yml publishes ghcr.io/mdopp/solaris-chat, ghcr.io/mdopp/solaris-gatekeeper (+ -ml), ghcr.io/mdopp/solaris-schema-init and ghcr.io/mdopp/solaris-wakeword-trainer on release tags (v*, via release-please) and pushes to main.

License

MIT. See LICENSE.