Tara LLM Config Node

Most of the Tara pack is about generating prompts. TaraLLMConfig is the node nobody gets excited about that makes the whole thing work, and it hides the pack's best feature in its default value: base_url ships as http://localhost:11434/v1. That's Ollama's OpenAI-compatible endpoint. In other words, this node is the bring-your-own-backend switch - point it at a local Ollama, LM Studio, llama.cpp, vLLM, or any OpenAI-compatible server, and the pack stops being a paid-API tool entirely. Free, local, private prompt rewriting. That's the case for this pack in one paragraph.

What it is

It's a config builder, not a caller. It packages ten settings into a single TARA_LLM_CONFIG object and outputs that on one wire; the actual API call happens downstream in TaraPrompterAdvanced or TaraAdvancedComposition, which construct the OpenAI client from this object. Nothing hits the network when this node runs - a detail worth knowing because it means you can build a whole config graph and queue it without spending a token.

The inputs, and which ones you actually touch:

• base_url - the endpoint. The Ollama default is the smart one; for Groq use https://api.groq.com/openai/v1, for OpenAI https://api.openai.com/v1.
• api_key - blank by default, which is fine for local servers that don't check keys (Ollama ignores it).
• llm_model - freeform text, default mixtral. Unlike the pack's preset dropdowns, this is not a fixed list, so you can type any model your endpoint serves - that's the escape hatch from the dated 2024-era choices.
• temperature (0.4), seed (42), max_tokens (1000), top_p (1), frequency_penalty / presence_penalty (0), timeout (60) - standard LLM sampling knobs, passed straight through. The seed is the notable one: on OpenAI/Groq it gives you repeatable rewrites for A/B testing guidance.

The single output, llm_config, is the TARA_LLM_CONFIG type. You can fan it out to multiple consumer nodes - one config driving several prompt or composition nodes is exactly how the pack expects to be used.

Where it sits

TaraLLMConfig → llm_config → TaraPrompterAdvanced (positive/negative)
                          └→ TaraAdvancedComposition (output_text)

Gotchas

• JSON mode is still forced downstream. TaraPrompterAdvanced sets response_format: json_object regardless of your config. Ollama's OpenAI shim handles it; a stripped-down server might 400. If you hit that, it's the endpoint, not the config node.
• The key can still leak to console. The consumer nodes inherited a debug print(api_key) from the legacy code. On a shared box, prefer the loader over pasting a key into this node.
• Local models on Ollama need the model pulled first - ollama pull mixtral (or whatever you name in llm_model) or you'll get a model-not-found error from the API.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ronniebasak/ComfyUI-Tara-LLM-Integration

Restart, or Manager → Install via Git URL. Requirements are just orjson and openai; no models to download. This is a small alpha from ronniebasak with barely any community footprint, so you're on the README and the source - but for a config node this simple, that's about all you need.