05 Gemini JSON

05 Gemini JSON sends a prompt to Gemini and gets back a JSON object shaped to your spec, instead of freeform prose. It's the node that turns an LLM from a text generator into something you can actually build a pipeline around - because JSON you can parse is data, and data can drive logic.

This is where ZMongo's two halves click together. Everything in this pack is built around structured documents in a database; a Gemini call that returns unstructured text is awkward to store. Gemini JSON's whole point is producing output that fits naturally into a ZMongo document - parameters, decisions, metadata - without you parsing a wall of prose.

How it works

Required inputs:

• session - from the API Key Session node.
• prompt - what you're asking for (default "Return a JSON object.").
• schema_json - the shape you want back, as a JSON schema. Defaults to {} (anything goes); the more you fill this in, the more reliably the model conforms. Give it fields, types, required keys - the backend passes the schema through to Gemini's structured-output handling.
• model - default gemini-2.5-flash.
• max_output_tokens - default 2048.
• temperature - default 0.2. For structured output, low is right; you want deterministic-ish, not creative.

Optional: system_instruction, gemini_prefix, refresh_token.

It POSTs to /gemini/api/json. Outputs:

• json - the full backend payload (for debugging).
• parsed_or_text - the useful one: the parsed JSON object serialized back to text if the model returned parseable structured output, or the raw text if it didn't. Either way you get something usable.
• success - boolean.

Notes

• Write a real schema. schema_json: {} is a "best effort" - Gemini may return JSON, but it won't be your JSON. A schema with explicit keys and types ({"type":"object","properties":{"positive":{"type":"string"},"cfg":{"type":"number"}}}) is the difference between data you can use and text you have to wrangle.
• Temperature 0.2 default is deliberate. Structured output and high temperature fight each other. If your JSON keeps coming back malformed, lower it, don't raise it.
• Because parsed_or_text is a plain STRING, you can drop it straight into a ZMongo document field or feed it to a JSON-parse node downstream.

Install

ComfyUI Manager → ComfyUI-ZMongo, or:

cd ComfyUI/custom_nodes
git clone https://github.com/CentralFloridaAttorney/ComfyUI-ZMongo

restart. Server-side Gemini call through the API session - thin local client, heavy pack requirements.txt on first import regardless.