DeepDoc: A Document Parsing Tool for Air-gapped RAG Built with Rust
Introducing DeepDoc, a text extraction tool for 13 document formats packed into a single binary using Rust, designed for offline RAG systems without cloud or Python dependencies.
Stock photo for illustration only, not from the actual event
- DeepDoc is a single-binary document text extraction tool supporting 13 formats, built with Rust.
- It operates 100% offline with zero cloud data uploads, no JVM required, and no ML model downloads.
- It accurately converts docx, pptx, xlsx, pdf, and other document families into Markdown.
- It focuses on preserving heading paths and data chunking optimized for RAG systems.
Building a RAG (Retrieval-Augmented Generation) system or document search feature often starts with the same old problem: converting docx, pdf, and pptx files into clean text, all under a strict constraint common in many organizations where documents must never leave the local machine.
Reviewing existing tools in the market reveals various distinct limitations:
- Apache Tika: A traditional standard that works well, but runs on the JVM, introducing memory management overhead and larger container image sizes.
- Python Stacks (e.g., Unstructured, Docling, MarkItDown): Popular in RAG tutorials, but require pip installation, download ML models, and run slowly for standard digital file processing.
- Cloud Parsers (e.g., LlamaParse): Convenient for getting started, but require uploading private data to external servers and charge per page, which fails to meet the needs of organizations like banks, hospitals, or law firms operating in air-gapped environments.
DeepDoc's approach was specifically developed to solve this problem by offering a single static Rust binary with no JVM, no Python, no model downloads, and zero network connections, delivering deterministic results every time the same dataset is inputted.
Modern RAG systems frequently struggle with the complexity of installing massive libraries and managing data security. Tools developed in Rust are emerging as an appealing choice for developers seeking speed, high performance, and complete control over data security directly on local machines.
In version v0.1, it supports the conversion of 13 born-digital document formats, including docx, pptx, xlsx, odt, ods, odp, epub, html, rtf, csv, md, txt, and text-based pdf, with CLI commands available to instantly scan folders recursively.
Stock illustration, not from an actual event
Regarding RAG implementation, standard chunking typically splits text by character count and causes a loss of context, but DeepDoc analyzes the document structure first and then splits it along block boundaries while retaining the heading path and byte range for each segment.
However, the developer clearly notes that DeepDoc v0.1 is not yet an OCR tool for parsing scanned image-based PDFs. If pure image files are encountered, the system will immediately return error code 4 to prevent garbage data from entering the RAG index, with OCR features planned as an optional add-on in the future.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment