| name | deploy-generator |
| description | Stand up the geniesim_generator scene-generation stack — the MCP asset servers + Open WebUI — via `docker compose`, picking one embedding backend. Trigger: When the user asks to "部署 generator", "deploy the scene generator", "启动资产检索服务", "start the MCP assets server", "run the generator stack", "set up open-webui for scene gen", or otherwise wants the generator's Docker services (`compose.yaml`, profiles `text` / `vl`) running.
|
| license | MPL-2.0 |
| metadata | {"author":"genie-sim","version":"1.0"} |
| prerequisites | [] |
| inputs | [{"name":"profile","desc":"Embedding backend profile (`text` or `vl`)","required":true},{"name":"host","desc":"Host to bind WebUI / MCP servers","required":false,"default":"0.0.0.0"}] |
| outputs | [{"desc":"Open WebUI + MCP assets servers running (default ports); `docker compose ps` shows them healthy"}] |
When to Use
- User wants the scene-generation backend running: the MCP asset/file servers
(gateway on
:8765) plus the Open WebUI front-end.
- User asks which embedding backend to pick, or how to configure the API key /
GPU model for asset retrieval.
Do not use for:
- Actually searching assets once the server is up →
search-assets skill.
- Driving the LLM to produce a scene →
generate-scene skill.
The ONE decision: which embedding backend
Asset RAG search needs an embedding backend. The two docker compose profiles
both bind the gateway to :8765, so only one runs at a time. Choose by
hardware / credentials:
| text profile | vl profile |
|---|
| Embedder | Qwen embedding API (Dashscope text-embedding-v4) | Qwen3-VL-Embedding local model |
| Hardware | No GPU | NVIDIA GPU + Container Toolkit |
| Credentials | Needs an API key | None (runs offline after weight download) |
| Modality | Text only | Image + text (better retrieval) |
| Extras | reranker N/A | optional Qwen3-VL reranker |
Ask the user (via AskUserQuestion) which they want if it isn't obvious from
context (do they have a GPU? do they have a Dashscope key?).
Workflow
All commands run from the generator package dir
(source/geniesim_generator/src/geniesim_generator/, where compose.yaml lives).
Prerequisites (both profiles)
The MCP servers import geniesim_assets (the multi-GB object library, shipped
as a separate package — not bundled in this image). It is now installed
on the host via pip install geniesim_assets; the compose stack mounts
the host's installed copy into each MCP container read-only (and
entrypoint.sh adds /opt to PYTHONPATH so the mount is importable).
