Siberia Ollama连接器 / Ollama Connector

The Siberia Ollama Connector is where every other node in this pack starts, and the name is a little misleading in the best way: it doesn't call any hosted API and needs no key. It's a handshake with your own Ollama server - the one you run on your machine or your LAN - and its job is to pick a server, pick a model, and hand a connection object to the pack's chat and vision nodes.

This is a mainstream pattern, not a novelty. Running a local LLM or vision model as a ComfyUI node and having it talk to an Ollama server is how thousands of people do prompt enhancement and captioning offline (the KB counts hundreds of Ollama-in-ComfyUI threads a year). What's less mainstream is this pack - it's small, recent, and has essentially zero community footprint, so treat the connection code as worth a read before you trust it. I did: it only talks to the URLs you put in its config file. No credential, no surprise phone-home.

How it works

The connector reads a config.yaml file inside its own folder, which defines your servers as name-plus-URL pairs:

ollama_servers:
  - name: Local Server / 本地服务器
    url: http://127.0.0.1:11434
  - name: Remote Server / 远程服务器
    url: http://your-remote-server:31434

Those names become the server_name dropdown. The model field is where the pack's cleverness lives: it starts as the literal string "刷新 / refresh", and a piece of JavaScript pokes the ComfyUI backend, which asks your Ollama server for its installed models and fills the dropdown with them. Pick "刷新 / refresh" and run, and the node auto-selects the first available model for you. There's also a timeout in seconds (5–300, default 30) for requests that hang.

The single output is 连接 / Connection, type OLLAMA_CONNECTION - a custom type that only the pack's own Siberia Ollama Chat and Siberia Ollama Vision nodes accept. It's the pack's private pipe: a dict carrying the server URL, the chosen model, the timeout, and the list of available models. You can't plug it into anything else, and nothing else can feed those two nodes. Set up the connector once, and it feeds both of them.

Getting it running

First make sure Ollama itself is alive and has models:

ollama serve
ollama pull qwen3-vl:8b   # or any model you actually want

Then install the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/siberiah2o/ComfyUI-SiberiaNodes.git
cd ComfyUI-SiberiaNodes
pip install -r requirements.txt

Restart ComfyUI. ComfyUI Manager users can search "ComfyUI-SiberiaNodes" instead of cloning by hand. Note the pack ships no model files - everything you run comes from Ollama, which is why the connector is the choke point.

Where people get burned

• The model dropdown only ever shows "刷新 / refresh". That means Ollama isn't reachable - server not running, or the URL in config.yaml is wrong. Test with curl http://127.0.0.1:11434/api/tags; you should get a JSON list of models back.
• The default remote URL is a placeholder. The shipped config.yaml points the remote server at 192.168.1.100:11434, which is almost certainly not yours. Edit the file to your real address or the connector will fail against a machine that doesn't exist.
• A model you pulled isn't in the dropdown. The list is fetched from the server when the frontend asks. If you pulled a model after the page loaded, re-select "刷新 / refresh" or reload the page.
• Models must already be pulled. The connector lists what Ollama has; it won't download anything for you.