Architecture Overview

Crate Dependency Graph

edgeparse-cli
  └── edgeparse-core
        └── pdf-cos

edgeparse-python
  └── edgeparse-core (via PyO3)

edgeparse-node
  └── edgeparse-core (via napi-rs)

Crates

Crate	Purpose
pdf-cos	Low-level COS object parser (xref, streams, filters)
edgeparse-core	Document model, extraction pipeline, renderers
edgeparse-cli	CLI binary
edgeparse-python	Python bindings (PyO3 + maturin)
edgeparse-node	Node.js bindings (napi-rs)

Pipeline Stages

EdgeParse processes a PDF through a 20-stage pipeline:

1. Byte loading — read file or buffer
2. Header parse — verify %PDF- magic
3. Xref parse — locate cross-reference table
4. Trailer parse — extract root, info, encrypt refs
5. Object resolve — dereference indirect objects
6. Stream decode — decompress FlateDecode, LZW, etc.
7. Page tree walk — enumerate pages from /Pages tree
8. Content stream parse — tokenize PDF operators (Tj, TJ, Tm, etc.)
9. Font map build — extract /ToUnicode, /Encoding, built-in maps
10. Glyph decode — map character codes → Unicode
11. Text run assembly — group characters into runs by font/position
12. Line detection — cluster runs into lines by baseline
13. Reading order — sort lines into column-aware reading order
14. Paragraph merge — join lines into paragraphs by indentation/spacing
15. Table detection — identify grid structures via ruling lines
16. Heading detection — classify headings by font-size heuristics
17. List detection — recognize bullet/numbered lists
18. Image extraction — decode inline/XObject images
19. Document model build — assemble Block tree
20. Rendering — output JSON, Markdown, HTML, or plain text

Design Principles

• Zero ML dependency — all extraction is rule-based
• Zero copy where possible — borrowed references into the PDF byte buffer
• No unsafe — safe Rust only (except FFI boundaries)
• Streaming capable — pages can be processed independently
• Deterministic — same input always produces same output