Skip to content
ironmark

AI pipelines

ironmark is built to be a reliable, deterministic parsing layer for AI agents and code generation tools: one parser, the same output for the same input, and machine-readable metadata about what it supports.

import { markdownToAst, markdownToHtml, astToMarkdown, CAPABILITIES } from "ironmark";
 
// 1. Inspect capabilities at startup
CAPABILITIES.formats; // ["html", "text", "ast", "markdown", "ansi", "jsx"]
 
// 2. Parse with deterministic options
const ast = markdownToAst(content, { preset: "llm" });
 
// 3. Render or transform
const html = markdownToHtml(content, { preset: "llm" });
const normalized = astToMarkdown(ast);

The llm preset

The llm preset produces deterministic, structure-first output:

  • Disables autolinks, wiki links, math and hard breaks, which are ambiguous in AI-generated text.
  • Enables heading ids and whitespace normalization.
  • Disables raw HTML passthrough, so it is safe by default.

Document structure

import { markdownToAst, extractHeadings, summarizeAst } from "ironmark";
 
const ast = markdownToAst(content, { preset: "llm" });
 
// Extract document structure
const headings = extractHeadings(ast);
// [{ level: 1, text: "Introduction", id: "introduction" }, ...]
 
// Summarize node types
const summary = summarizeAst(ast);
// { blockCount: 5, nodeCounts: { Heading: 2, Paragraph: 3, ... } }

For chunking and embeddings, markdownToText() gives the words without markup; see Plain text. A document returns text, headings and links from one parse.

Edit this page on GitHub