compile

name	compile
description	Use when: building or updating wiki knowledge from markdown sources. Handles both quick compile (single source) and deep compile (cross-source synthesis). Do NOT use when: importing new material (use ingest) or health-checking (use lint).

Compile

Transform markdown extractions into structured wiki knowledge.

Quick Compile（单源编译）

Triggered after ingest. Input: one markdown file from markdown/. Output: one summary in wiki/sources/.

Output Requirements

1. YAML frontmatter — must include:

   • title: article title
   • source: original URL or file path (from input frontmatter)
   • tags: 3-8 topic tags (lowercase, hyphenated)
   • date: ingest date (from input frontmatter)

2. Summary body — Obsidian-compatible markdown:

   • Concise but comprehensive summary of the article
   • Key takeaways as a bullet list
   • Use [[wikilinks]] to link related concepts (even if the target doesn't exist yet)
   • Preserve source attribution — never fabricate claims not in the original

3. Language — match the language of the source material

Example Output

---
title: Attention Is All You Need
source: https://arxiv.org/abs/1706.03762
tags: [transformer, attention, deep-learning, nlp]
date: 2026-04-03
---

## Summary

The paper introduces the Transformer architecture...

## Key Takeaways

- Self-attention replaces recurrence and convolution
- Multi-head attention allows attending to different representation subspaces
- Positional encoding compensates for the lack of recurrence

## Related

- [[attention-mechanism]]
- [[sequence-to-sequence]]

Deep Compile（深度编译）

Cross-source synthesis. You have full write access to the wiki — create, update, and reorganize as needed.

What You Do

1. Read _index/master.md to understand current state
2. Identify new/unprocessed sources — sources in sources/ that haven't been synthesized into concepts yet (check wikilink references)
3. Create or update concepts/ articles — synthesize knowledge across multiple sources into standalone concept articles
4. Create or update maps/ articles — high-level overviews, timelines, comparison tables that connect multiple concepts
5. Update _index/ — keep master.md, by-topic.md, and other indexes current
6. Update SCHEMA.md — if the wiki's organizational structure has evolved, reflect it

Concepts (concepts/)

A concept article synthesizes knowledge about one topic from multiple sources.

---
title: Transformer Architecture
tags: [transformer, deep-learning, architecture]
sources: [attention-is-all-you-need, bert-paper, gpt-overview]
date: 2026-04-03
---

## Overview

Brief definition and significance.

## Key Ideas

- Idea 1 — synthesized from [[sources/attention-is-all-you-need]] and [[sources/bert-paper]]
- Idea 2 — ...

## Related Concepts

- [[self-attention]]
- [[positional-encoding]]

Rules:

• A concept must reference at least 2 sources — single-source knowledge stays in sources/
• Always cite which sources support each claim
• Use [[wikilinks]] liberally to connect concepts

Maps (maps/)

A map provides a bird's-eye view across multiple concepts.

Types:

• Topic overview — "Map of Deep Learning Architectures"
• Timeline — chronological progression of a field
• Comparison — structured comparison table across approaches
• Learning path — suggested reading order for a topic

---
title: Deep Learning Architectures Overview
type: topic-overview
tags: [deep-learning, architecture, map]
date: 2026-04-03
---

## Landscape

Brief overview of the field.

## Architectures

| Architecture | Key Innovation | Era | Sources |
|---|---|---|---|
| CNN | Spatial feature extraction | 2012 | [[sources/alexnet]] |
| Transformer | Self-attention | 2017 | [[sources/attention-is-all-you-need]] |

## Related Concepts

- [[concepts/transformer-architecture]]
- [[concepts/convolutional-networks]]

Rules:

• Maps synthesize across concepts, not directly from sources
• Always include links to the concepts and sources that inform the map

Incremental Strategy

• Only process sources that are new since last deep compile
• When updating existing concepts/maps, extend rather than rewrite — add new information, update dates
• If a concept has grown too large, consider splitting it
• If new sources reveal connections between existing concepts, create a new map

SCHEMA.md Maintenance

After making structural changes, update SCHEMA.md to reflect:

• New topic categories that have emerged
• Changes to the organizational hierarchy
• New conventions or patterns you've established