EmbedM

version 1.0.1

A Markdown compiler driven by source files.

• Project Background
• How It Works
• Use Cases
  • Keeping code documentation in sync
  • Live metadata in a README or changelog
  • Data tables without copy-paste
  • CI drift detection
  • AI agent context documents
• Directives
• Quick Start
• Features
• Documentation
• License
• Contributing

Project Background

EmbedM is part of an exploration into how far AI-assisted development can go when building a non-trivial tool that could be used in a production CD/CI chain. This project has been built based on a human defined architecture, functional spec and a series of interface contracts, then implemented using using Claude and to a lesser extent Google Gemini.

How It Works

EmbedM compiles Markdown documents from directive blocks. Each directive references a source — a code file, a data query, a CSV table, or another document — and is replaced with the extracted, formatted content on compile. Change the source; recompile; the document is current.

Use Cases

Keeping code documentation in sync

Embed a function directly from the source file, scoped by a named region or by symbol name. When the implementation changes the docs regenerate on the next compile — no copy-paste, no drift. Instead of copying the function code, you simply add a reference to the class/function/method/enum or struct.

Instead of adding code that may go out of date:

public void createUser(string user) {
    // ...
}

You create a link to said method, which will be replaced with the up-to-date function at compile time, or give a clear error in case the method 'createUser' is no longer there.

type: file
source: src/api/handlers.java
symbol: UserHandler.createUser
title: "POST /users"
link: true

Live metadata in a README or changelog

Pull version numbers, project names, and other values from pyproject.toml, package.json, or any JSON/YAML/TOML/XML file. The version at the top of this page is a live example — it is compiled from pyproject.toml at build time. Instead of a hard coded version, create a reference to the project. Eg:

type: query-path
source: pyproject.toml
path: project.version
format: "Released: **v{value}**"

Data tables without copy-paste

Embed CSV, TSV data or structured json as formatted Markdown tables. Apply column selection, filtering, and sorting inline — the source file is the single source of truth via:

type: table
source: reports/q4-summary.csv
select: "Region as Region, Revenue as Revenue_USD"
order_by: "Revenue_USD desc"
limit: 10

CI drift detection

Use --verify in your pipeline to catch documentation that has fallen behind its sources. Exit code 1 if any compiled file is stale.

embedm ./docs/src --verify -d ./docs/compiled

AI agent context documents

Use recall to query a large document — a devlog, a decision log, an ADR set — and extract the sentences most relevant to a given topic. Compose multiple queries into a single compiled context file that an AI assistant reads at session start.

type: recall
source: ./devlog.md
query: "validation transform boundary error handling"
max_sentences: 5

EmbedM itself uses this: its agent context file is compiled from the project devlog using four targeted recall queries — plugin conventions, architectural rules, common mistakes, and the active spec. The context window stays focused without manual curation.

Directives

Directives are fenced YAML blocks tagged ```yaml embedm. On compile, each is replaced in-place with the extracted content:

type: file
source: src/config/defaults.py
region: connection_defaults

# connection_defaults
HOST = "localhost"
PORT = 5432
TIMEOUT = 30
POOL_SIZE = 10

Structured data queries render inline:

type: query-path
source: config/app.yaml
path: database.pool_size
format: "Default pool size: **{value}**"

Default pool size: 10

Quick Start

Install

pip install embedm

Or from source:

git clone https://github.com/Fultslop/embedm.git
cd embedm
pip install -e .

Compile a single file

embedm content.md -o compiled/content.md

Compile a directory

embedm ./docs/src -d ./docs/compiled

Preview without writing

embedm content.md -n

Check that compiled files are up to date

embedm ./docs/src --verify -d ./docs/compiled

Generate a default config file

embedm --init

Creating new plugins

See the plugin_tutorial

Features

File embedding

• Embed entire files, line ranges (5..10), or named regions (md.start:name / md.end:name)
• Markdown sources are merged inline; all other types are wrapped in a fenced code block
• Optional title, source link, and line-number annotation

Symbol extraction

• Extract classes and methods by name from C/C++, C#, and Java source files
• Dot-notation for nested symbols: OuterClass.InnerClass.methodName
• Overload disambiguation: add(int, int) vs add(int, int, int)

Structured data

• Query any value from JSON, YAML, TOML, or XML using dot-notation paths
• Scalars render inline; dicts and lists render as YAML code blocks
• Format strings for inline interpolation: "version {value}"

Data tables

• Render CSV and TSV files as Markdown tables
• Column selection, row filtering (exact match and comparison operators), sorting, pagination

Table of contents

• Auto-generated from document headings, including headings in embedded files
• GitHub-compatible anchor links

AI context

• synopsis — generate a condensed summary of a document
• recall — build structured retrieval blocks for AI agent context files

Recursive embedding

• Markdown files that embed other Markdown files, up to a configurable depth

Safety

• Configurable limits on file size, memory, recursion depth, and embed output size
• --verify mode for CI drift detection

Documentation

License

MIT License — see LICENSE file for details.

Contributing

Contributions are welcome. Please open an issue to discuss proposed changes before submitting a pull request.