README.md

metadata

sdk: gradio

LLM LSP

A proof of concept for intelligent document editing that works like humans do - scan, jump, edit.

The Idea

Instead of making AI read entire documents, we let it:

1. Scan headers to find location (like grep)
2. Jump to right section (like sed)
3. Make targeted changes
4. Put it back (like cat)

Quick Start

pip install -r requirements.txt  # openai, gradio

from lsp import LSP
editor = LSP(api_key="your-key")
editor.edit("doc.md", "add version warning")

How It Works

1. Map Sections Fast

sections = find_sections("doc.md")  # Uses grep-like scan

2. AI Picks Target

{
    "sections": {"## Setup": 10, "## Config": 50},
    "task": "add version warning",
    "target": "## Setup at line 10"
}

3. Extract & Edit

content = extract_section(file, 10, 49)  # Just that piece
new_content = ai.modify(content)         # AI edits small part

4. Put It Back

replace_section(file, 10, 49, new_content)

Try It

python main.py  # Launches Gradio UI

Status

• POC stage
• Works with markdown files
• Uses gpt-4o-mini
• Basic UI

Limits

• One section at a time
• Just markdown for now
• No fancy error handling
• Will fuck up if sections aren't clear

Next Steps?

• Handle multiple sections
• Better section detection
• More formats
• Whatever you need it to do