llm-markdown-sanitizer (Python)

Fix broken markdown that LLMs generate — tables, lists, headings, emphasis, code fences, quotes. Zero dependencies, one function.

A Java binding with the same behavior is also available — see the repository root for both.

Install

PyPI page: https://pypi.org/project/llm-markdown-sanitizer/ (release history, file hashes, full metadata).

Requires Python 3.9+. No other dependencies get pulled in.

pip install llm-markdown-sanitizer

Using uv (if that's your workflow):

uv add llm-markdown-sanitizer

Using a virtual environment (recommended for any real project):

python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install llm-markdown-sanitizer

Pin a specific version if you want reproducible builds — see the PyPI release history for available versions:

pip install "llm-markdown-sanitizer==0.2.1"

Add it to requirements.txt / pyproject.toml the normal way:

llm-markdown-sanitizer>=0.2.1

Verify it installed correctly:

python -c "from llm_markdown_sanitizer import clean_markdown; print(clean_markdown('**hi**there'))"
# **hi** there

Use

The whole API is one function:

from llm_markdown_sanitizer import clean_markdown

clean_markdown("**Note**this needs a space")
# "**Note** this needs a space"

clean_markdown("| A | B | | --- | --- | | 1 | 2 |")
# "| A | B |\n| --- | --- |\n| 1 | 2 |"

Default settings handle the common failure modes without additional configuration.

In a FastAPI endpoint

A typical place to call this is right before a stored or freshly-generated LLM response goes out to a client:

from fastapi import FastAPI
from llm_markdown_sanitizer import clean_markdown

app = FastAPI()

@app.get("/lectures/{lecture_id}/summary")
def get_summary(lecture_id: int):
    raw = db.get_ai_summary(lecture_id)  # however you fetch/generate it
    return {"summary": clean_markdown(raw)}

Streaming/multi-part LLM responses

Some SDKs return responses as a list of {"text": ...}-shaped chunks instead of one string. clean_markdown accepts that directly:

chunks = [{"text": "# Hello"}, {"text": "\n\nWorld"}]
clean_markdown(chunks)
# "# Hello\n\nWorld"

Why this exists

Ask an LLM to answer in markdown and eventually you'll get: the whole answer wrapped in a stray ```markdown fence, **bold**text glued directly onto the next word, list indentation that's inconsistent within the same response, and tables that are either collapsed onto one line or missing a separator row. Rendering that output as-is breaks the UI.

clean_markdown() fixes all of the above in a single left-to-right pass over the text — no whole-string regex backtracking, so it stays fast on long documents.

What it fixes

Protecting your own syntax

If your prompts produce custom tokens — a [[wiki]]-style syntax, template placeholders, etc. — that the cleanup passes above might mangle, they can be excluded explicitly:

import re

clean_markdown(text, protect_patterns=[re.compile(r"\[\[.*?\]\]")])

Origin

Extracted from the markdown-cleanup layer of a production RAG service, after months of hardening against real LLM output. The domain-specific parts — a custom wiki syntax, a Korean-language note pattern — were removed in favor of the general protect_patterns mechanism above, so callers can supply their own domain syntax instead.

Contributing

Bug fixes and small improvements are welcome. No CLA/DCO required — see CONTRIBUTING.md for guidelines and how to run the test suite locally. AI coding agents should pick up AGENTS.md automatically.

License

MIT