(Deprecated) Tara LLM Primary Node

TaraPrompter is the node that started this pack, and the README now politely calls it "(Deprecated) Tara LLM Primary Node". Deprecated doesn't mean useless - it means the pack outgrew it. It's still the cleanest demonstration of what the entire Tara project does: hand a large language model your rough positive and negative prompts plus some guidance, and get back a refined positive/negative pair ready to feed into a CLIP text encoder. If you're trying to figure out whether this pack is for you, this node is the 60-second answer.

How it works

The mechanism is a chat completion with a hard requirement baked in: the model must reply in JSON. A system message (the pack's META_PROMPT) demands the format {"positive": "...", "negative": "..."}, the request goes out with response_format set to json_object, and the reply is parsed with orjson and then scrubbed - braces, brackets and quotes stripped - before the two strings are returned as positive and negative outputs. Those wire straight into CLIPTextEncode. That's the whole loop.

The inputs:

• api_key and model - the key, and a dropdown of the pack's 2024-era model list. groq/mixtral-8x7b-32768 is the one the README pushes, and the free-Groq story was the pack's real selling point: an LLM that costs nothing to run.
• guidance - multiline instructions for how to refine. This is where you say "keep it under 40 tokens, no quality spam, add camera language".
• prompt_positive / prompt_negative - the raw material.

Why it's deprecated

The design was rigid. The model list is hardcoded and dated (gpt-4-turbo-preview era), the endpoint is hardcoded per provider (OpenAI and Groq only), and there's no temperature, seed, or token control from the UI. Two real annoyances on top: it needs json_object response format, so any OpenAI-compatible endpoint that doesn't support JSON mode will just 400 you, and the code prints your API key to the ComfyUI console log on every run - a debug line that never got cleaned up. On a shared box that's a leak you didn't ask for.

The replacement is TaraPrompterAdvanced, which takes a llm_config object instead of raw inputs and inherits everything else. Wire old workflows onto the new node and you get seed support, configurable sampling params, and any OpenAI-compatible backend.

Install and gotchas

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

Deps are orjson and openai - nothing else. No model downloads. The common failure is the JSON-mode 400 with self-hosted or quirky endpoints; if you see that, the config-based nodes (TaraLLMConfig pointing at a compatible server) are the escape hatch, and honestly the one you want anyway. The pack itself is a small alpha from ronniebasak that's been quiet since 2024 - treat it as a tool, not a maintained product, and expect the deprecated nodes to keep working but never improve.