Civitai Chat Completion

The simple chat node in this pack is a wrapper; Civitai Chat Completion is the actual engine. It's an OpenAI-compatible chat-completion call with the whole parameter surface exposed - temperature, penalties, tools, structured JSON output - and it's the node you reach for when a single prompt-in/prompt-out wrapper isn't enough.

What it is

A chat completion step on Civitai's orchestration API. You hand it a model name and a messages_json array (the standard [{role, content}, ...] shape), and it returns the response plus every OpenAI field you'd expect. No local model, no VRAM - every call is billed in Buzz, same as the rest of the pack.

Inputs that matter

Required: model (e.g. "gpt-4o", "gpt-4-turbo", "gpt-4o-mini") and messages_json - a JSON array of messages, as text. This is one of the pack's "freeform structures stay JSON text" inputs, so you paste or compute the JSON string; the node validates it before submitting and errors locally if it doesn't parse.

The optional list is long but only a few earn your attention:

• max_tokens (default 1 - remember to raise it), temperature (0–2), top_p - the usual sampling controls.
• tools_json and tool_choice_json - define tools the model can call; tool_choice can be "auto", "none", "required", or an object naming a specific function.
• response_format_json - when set to json_object or json_schema, the emitted content is parsed server-side and surfaced on the parsed output, which downstream steps can reference field-by-field (e.g. output.parsed.<field> via $ref). This is the structured-output path for building agents that produce data, not prose.
• modalities_json - default text-only; include "image" and the request routes to the image generation pipeline, returning generated images on the assistant message. That's the wildcard: one node that can hand back pictures.
• n - how many completions to generate.

Outputs

Ten of them, and the useful ones are: choices (the completions), usage (token counts - watch this on your Buzz bill), parsed (the structured-JSON result), and model/id/created for bookkeeping. Plus the pack-standard workflow_id and raw_json.

Install & gotchas

Same pack install as everything here - ComfyUI Manager (search Civitai Comfy Nodes), comfy node registry-install civitai-comfy-nodes, or git clone + pip install -r requirements.txt. Auth via CIVITAI_API_TOKEN, stored OAuth, or a wired Civitai Auth node.

Where beginners trip:

1. The default max_tokens is 1. The schema default really is 1, so a fresh node returns a single token until you raise it. Bump to a few hundred.
2. messages_json must be valid JSON or the node fails fast with "Input 'messages_json' is not valid JSON". Single quotes will bite you - it needs proper JSON quoting.
3. Cost visibility. Every call is metered. raw_json includes usage/cost info; get in the habit of reading it on expensive runs. A 402 means out of Buzz.
4. It's a hosted frontier model, so filtering follows the provider. Unlike a local abliterated LLM, there are no weights to patch - whatever the service refuses, the node refuses.

Honest take: for a one-off text job, Civitai Chat (Simple) is friendlier. This node earns its complexity when you need tools, JSON output, or multimodal results inside a real workflow - that's where having an OpenAI-grade chat API as a graph node actually shines.