Siberia Ollama聊天 / Ollama Chat

The Siberia Ollama Chat node is what happens when you want a real text conversation with a local model inside your graph - not a one-shot prompt rewrite, but an actual back-and-forth where the model remembers what you said before. You've seen this job done by a thousand standalone LLM tools; here it's a node you wire into a workflow, fed by the pack's connector and read out through its display node.

It's part of the same family as the pack's vision node, but with the camera turned off: pure text in, text out, running entirely on your own machine through Ollama. Offline, free per call, uncensored if you pulled an uncensored model - the reasons the KB keeps pointing people at the local path for this kind of work.

What you set

The required inputs are message (a multiline text box - your question) and clear_history (a boolean, default false). Everything else is optional:

• connection - the OLLAMA_CONNECTION from the Siberia Ollama Connector. Leave it unconnected and the node falls back to defaults and talks to your local server anyway.
• temperature - 0.1 to 1.0, default 0.7. Lower is more deterministic and repetitive; higher is looser. For rewriting prompts or captions you'll usually want the low end.
• max_tokens - 1024 to 32768, default 4096. One honest warning: the schema lets you dial this up to 32768, but the SDK client underneath clamps generation to 8192 tokens no matter what you ask for. Setting 32768 just sets a number that gets ignored past 8192.
• language - 中文 or English. This isn't translation; it sets the system prompt, so the model is instructed to answer in that language.

The single output is 回复 / Response (STRING) - the model's reply text. Wire it to the pack's Universal Display node to actually read it on the canvas.

The history thing

The gotcha that bites everyone: the node keeps a chat_history list on itself, and it accumulates across runs. Every time you queue, your new message is appended to everything before it, so the model genuinely remembers the conversation. That's the feature - until it isn't. A workflow that's been through a dozen runs carries a dozen turns of context, the token budget quietly inflates, and answers start drifting toward whatever the conversation was about an hour ago. Flipping clear_history to true on a run wipes the slate. If your replies suddenly read like they're responding to a conversation you don't remember having, this is why.

Install and wiring

Same pack, same steps:

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

Restart ComfyUI and it's under Siberia Nodes/Ollama. Manager users search "ComfyUI-SiberiaNodes". You need Ollama running (ollama serve) with at least one text model pulled - a chat model like a Llama or Qwen GGUF via ollama pull, not necessarily a vision one.

The intended shape is the pack's example workflow: Connector → Chat → Universal Display. One honest note: this is a small, near-unknown pack, and an LLM chat node is exactly the category that's been weaponized once before in this ecosystem (ComfyUI_LLMVISION). I read the chat code and it just talks to your Ollama server - but for a node this obscure, reading before trusting is the right habit regardless.

Troubleshooting

• Reply is an "Error: …" string rather than text. The node returns errors as normal output instead of crashing, so read the actual string. Most common: Ollama unreachable, or no models on the server.
• Replies stopped making sense. Check the accumulated history; run once with clear_history on.
• Language ignored. If the model stubbornly answers in the wrong language, it's usually a weak model, not a bug - the system prompt is an instruction, not a guarantee.