Markdown doc. generator

The name is the whole pitch: this node takes a chunk of source code and gets a local LLM to document it in Markdown. What's notable - and increasingly rare in the LLM-node space - is that the name isn't lying about needing a key. There's no API key, no cloud account, no billing page. It talks to an OpenAI-compatible endpoint on your own machine, defaulting to http://localhost:5001/v1/chat/completions, which is exactly the endpoint Koboldcpp exposes out of the box. That's the pairing the author tested the LLM nodes against.

How it works: the prompt input is the source file to document - paste in a Python file or a section of it. temperature (default 0.5) controls randomness, max_tokens (default 2000) caps how much the model writes, seed (default 42) keeps the output reproducible, and url points at your endpoint. The optional extra_context input lets you hand the LLM background material - "out of scope constants and helper definitions" - so it can explain code that references things outside the snippet. Under the hood it builds a chat-completions request, fires it at the endpoint, and returns four outputs: request_json and response_json (the raw request/response, useful when you're debugging why the model answered oddly) and markdown plus markdown_list (the generated doc, single string and as a list of lines).

The intended pipeline is very LF: use LF_RegionExtractor to pull the code regions you care about out of a folder of files, feed them into this node, then pass markdown to LF_SaveMarkdown to write the doc to disk. The nodes even share a category story - "LF Nodes/LLM" pairs with the IO and save nodes to make a mini documentation pipeline that runs entirely offline.

The setup cost is the LLM server, not the node. You need something serving an OpenAI-compatible API locally - Koboldcpp is the tested default, and any unauthenticated OpenAI-compatible endpoint should work per the README. The author's sample setup uses UCLA-AGI/Llama-3-Instruct-8B-SPPO-Iter3 with a LLaVA mmproj for vision, but any decent instruct model will produce usable docs. Realistically you want at least a 7B/8B model on a GPU; a 3B will write docs that read like a tired intern.

Where people get burned, in order:

• You forgot the server. The node errors if nothing is listening on port 5001. Start Koboldcpp first, then run the workflow.
• The default port isn't yours. If your server is elsewhere, update url - this is the field everyone's first error message points at.
• Everything on the graph is a dependency now. Like all LLM-node workflows, this one is only as deterministic as the model you run; seed helps, but a different model version can produce different docs.

Install the pack via ComfyUI Manager (search "LF Nodes") or git clone https://github.com/lucafoscili/comfyui-lf into custom_nodes, then restart. And the pack reminder that's worth repeating everywhere: comfyui-lf is legacy/frozen as of early 2025 - fully working, but new work happens in lucafoscili/lf-nodes.