pixrep 0.1.7

Convert code repositories into structured PDF collections for LLM collaboration.

pixrep

📉 SAVE UP TO 90% TOKENS

Turn Codebases into Visual Context for Multimodal LLMs

According to DeepSeek-OCR research and local benchmarking, visual encoding (PDF) outperforms plain-text ingestion for massive repositories.

---

📖 Introduction

pixrep is a developer tool designed to bridge the gap between large code repositories and Multimodal Large Language Models.

Instead of feeding raw text that consumes massive context windows, pixrep converts your repository into a structured, hierarchical set of PDFs. This allows you to:

• Save 90% Tokens: Visual encoding is far more efficient than text tokenization.
• Test for Free: Easily share your entire codebase with premium models (like Claude Opus 4.6) on platforms like arena.ai without hitting text limits.

🚀 Why Visual Code? (The 90% Claim)

Traditional RAG (Retrieval-Augmented Generation) relies on raw text. However, recent research (including the DeepSeek-OCR paper) indicates that visual encoders can represent dense information more efficiently than textual tokenizers.

• Text Tokenization: 1 page of dense code ≈ 500-800 text tokens.
• Visual Tokenization: 1 page of code (PDF image) ≈ Fixed patch count (e.g., 85-256 tokens depending on the model).

pixrep creates a layered PDF structure:

1. Macro View (00_INDEX.pdf): A visual map of the directory tree and project statistics.
2. Micro View (File PDFs): Syntax-highlighted, line-numbered renderings of individual code files.

This approach enables an Agentic workflow: Read the Index -> Identify relevant files -> Ingest only specific PDFs.

✨ Features

• 📉 High Efficiency: Drastically reduces context window usage for large repos.
• ⚡ Faster Scanning: Single-pass file loading (binary check + line count + optional content decode) to reduce I/O overhead.
• 🎨 Syntax Highlighting: Supports 50+ languages (Python, JS, Rust, Go, C++, etc.) with a "One Dark" inspired theme.
• 🧠 Semantic Minimap: Auto-generates per-file micro UML / call graph summaries to expose structure at a glance.
• 🔥 Linter Heatmap: Integrates ruff / eslint findings and marks risky lines with red/yellow visual overlays.
• 🗂️ Hierarchical Output: Generates a clean 00_INDEX.pdf summary and separate files for granular access.
• 🌏 CJK Support: Built-in font fallback for Chinese/Japanese/Korean characters (Auto-detects OS fonts).
• 🛡️ Smart Filtering: Respects .gitignore patterns and supports custom ignore rules.
• 📊 Insightful Stats: Calculates line counts and language distribution automatically.
• 🧾 Scan Diagnostics: Prints scan summary (seen/loaded/ignored/binary/errors) for faster troubleshooting.

📦 Installation

pip install pixrep

🛠️ Usage

Quick Start

Convert the current directory to PDFs in the default output folder (./pixrep_output/<repo_name>):

pixrep .

Common Commands

Generate PDFs for a specific repo:

pixrep generate /path/to/my-project -o ./my-project-pdfs

Pack core code into a single minimized PDF (all-in-one):

pixrep onepdf /path/to/my-project -o ./ONEPDF_CORE.pdf

Notes:

• Defaults to git ls-files (tracked files) when available.
• Defaults to "core-only" filtering (skips docs/tests); use --no-core-only to include them.

Preview structure and stats (without generating PDFs):

pixrep list /path/to/my-project

list mode now uses lightweight scanning (no file content decode), so large repos respond significantly faster.

Show only top 5 languages in the summary:

pixrep list . --top-languages 5

CLI Reference

Argument	Description	Default
repo	Path to the code repository.	. (Current Dir)
-o, --output	Directory to save the generated PDFs.	./pixrep_output/<repo>
--max-size	Max file size to process (in KB). Files larger than this are skipped.	512 KB
--ignore	Additional glob patterns to ignore (e.g., *.json test/*).	[]
--index-only	Generate only the 00_INDEX.pdf (Directory tree & stats).	False
--disable-semantic-minimap	Turn off per-file semantic UML/callgraph panel.	False
--disable-lint-heatmap	Turn off linter-based line heatmap background.	False
--linter-timeout	Timeout seconds for each linter command.	20
--list-only	Print the directory tree and stats to console, then exit.	False
-V, --version	Show version information.	-

⚙️ Performance Notes

pixrep now applies two execution paths:

1. Light scan path (pixrep list, pixrep generate --index-only, --list-only): only metadata and line counts are collected; file content is not loaded.
2. Full scan path (regular pixrep generate): file content is decoded only when needed for PDF rendering.

This reduces memory pressure and disk I/O for repository exploration workflows.

📂 Output Structure

After running pixrep ., you will get a folder structure optimized for LLM upload:

pixrep_output/pixrep/
├── 00_INDEX.pdf             # <--- Upload this first! Contains tree & stats
├── 001_LICENSE.pdf
├── 002_README.md.pdf
├── 003_pixrep___init__.py.pdf
├── 005_pixrep_cli.py.pdf
└── ...

🧩 Supported Languages

pixrep automatically detects and highlights syntax for:

• Core: Python, C, C++, Java, Rust, Go
• Web: HTML, CSS, JavaScript, TypeScript, Vue, Svelte
• Config: JSON, YAML, TOML, XML, Dockerfile, Ini
• Scripting: Bash, Lua, Perl, Ruby, PHP
• And more: Swift, Kotlin, Scala, Haskell, OCaml, etc.

🤝 Contributing

We welcome contributions! Please feel free to submit a Pull Request.

1. Fork the repository.
2. Create your feature branch (git checkout -b feature/AmazingFeature).
3. Commit your changes (git commit -m 'Add some AmazingFeature').
4. Push to the branch (git push origin feature/AmazingFeature).
5. Open a Pull Request.

📄 License

Distributed under the MIT License. See LICENSE for more information.