Ollama Chat

OllamaGenerateV2 handles a single ask-and-answer well. OllamaChat is for when you actually want a conversation - several turns, with the model remembering what was said earlier, built on Ollama's dedicated chat endpoint (ollama.chat) rather than the plain generate call. The pack's own description is direct about the split: this one is "designed specifically for multi-turn conversations and cloud models with your Ollama public key," where GenerateV2 is closer to a single structured request.

How it works

The mechanical difference from GenerateV2 is where conversation state lives. GenerateV2 round-trips a context blob through your graph manually. OllamaChat keeps history natively inside the node and hands you a dedicated history socket (OLLAMA_HISTORY type) instead - wire one Chat node's history output into the next one's history input and you've got a continuing thread, or split it to branch two different follow-ups off the same conversation point. reset_session clears the current history on demand. Worth repeating the node's own warning here: if multiple nodes share the same history ID, resetting from any one of them clears it for all of them - not just the node you clicked reset on.

Inputs and outputs that matter

• system and prompt - same shape as GenerateV2: system sets role/behavior, prompt is the actual turn.
• think - enables a reasoning pass before answering; only works on models trained to support it, fails outright on ones that aren't.
• format - text or json.
• connectivity (optional) - an OllamaConnectivityV2 node; required unless you're supplying meta instead. Same hard rule as GenerateV2: leave both empty and the node can't run.
• options (optional) - an OllamaOptionsV2 bundle.
• images (optional) - for vision-in-chat; make sure the model actually supports it, or you'll get a confident, wrong answer instead of an error.
• meta (optional) - chains connectivity/options from another Chat node, same pattern as GenerateV2.
• history (optional, OLLAMA_HISTORY) and reset_session (bool) - the conversation state and the clear switch.
• Outputs: result (the answer), thinking, meta, and history to pass forward.

Installing it

Via ComfyUI Manager: search comfyui-ollama (by Stav Sapir), or ollama → stavsap. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/stavsap/comfyui-ollama
pip install -r requirements.txt

then restart ComfyUI. If you're pointing this at an Ollama Cloud model that requires auth, set that up on the Ollama side, not on this node - either run ollama signin, or manually add your public key at ollama.com/settings/keys. The key file itself lives at a fixed path per OS (~/.ollama/id_ed25519.pub on macOS, for example), nothing you paste into ComfyUI directly.

Common issues

Node won't run at all. Check that connectivity or meta is actually plugged in - same requirement as GenerateV2, and just as easy to forget when you're mid-rewire.

think breaks generation. Same story as GenerateV2's version of this field: not every model supports a reasoning step, and forcing it on an unsupported one fails rather than degrading gracefully.

Resetting one node wipes a shared history you didn't mean to touch. If you're building a graph with branching conversations off a shared history ID, reset_session isn't node-local - it clears the whole history everywhere it's used. Give each branch its own history if you need independent resets.

Not sure whether you need this over GenerateV2. If you're doing a single ask, or already round-tripping context manually and it's working, there's no reason to switch. Reach for Chat specifically when you want native multi-turn state without wiring context by hand.