LiteParse Document Parsing Library

LiteParse is a standalone, open-source PDF parsing tool designed for fast and lightweight local parsing. It focuses on high-quality spatial text parsing with bounding boxes and does not require proprietary LLMs or cloud dependencies.

Key Features:

• Fast Text Parsing: Uses PDFium for spatial text extraction.
• Flexible OCR: Supports built-in Tesseract (zero setup) or external HTTP OCR servers (e.g., EasyOCR, PaddleOCR).
• Complexity Detection: Ability to check if a document requires OCR before full processing.
• Multiple Output Formats: Supports Markdown (with headings, tables, lists, images, and links), JSON (with bounding boxes), and plain text.
• Screenshot Generation: Creates high-quality page screenshots for LLM agents.
• Multi-language/Platform Support: Available for Rust, Node.js/TypeScript, Python, and WASM (Browser) across Linux, macOS, and Windows.

What works in the browser

• PDF parsing using Uint8Array input.
• Custom OCR via the ocrEngine callback interface.
• Output formats: Text, JSON, and Markdown.
• Complexity detection: Using parser.isComplex(bytes).
• Full extraction options: Annotations, form fields, structure trees, vector graphics, etc.

What doesn't work

• File path input: You must pass Uint8Array instead of a string path.
• Office conversion: DOCX/XLSX/PPTX conversion is unavailable (requires LibreOffice).
• Built-in OCR: Tesseract and HTTP OCR backends are not available; you must provide a custom ocrEngine.
• Screenshots: Not supported in the WASM build.
• Multi-threading: numWorkers is not available; parsing is single-threaded.
• Filesystem access: imageOutputDir is unavailable; use extractImages and read the bytes from the result instead.

LiteParse supports multiple OCR strategies:

1. Built-in Tesseract

Tesseract is included out-of-the-box. You can specify languages or provide a custom path to trained data.

• Specify language: --ocr-language <lang> (e.g., fra)
• Custom data path: --tessdata-path <path> or via environment variable TESSDATA_PREFIX.

2. HTTP OCR Service

You can connect to an external OCR service (like EasyOCR or PaddleOCR) by providing a URL.

• Server URL: --ocr-server-url <url>

Custom OCR API Requirements: To use a custom service, implement an endpoint that:

• Responds to POST /ocr with file and language parameters.
• Returns JSON: { "results": [{ "text": "...", "bbox": [x1,y1,x2,y2], "confidence": 0.0 }] }

lit parse document.pdf --ocr-language fra
export TESSDATA_PREFIX=/path/to/tessdata
lit parse document.pdf

LiteParse supports automatic conversion of various document formats to PDF before parsing.

Office Documents

Conversion requires LibreOffice to be installed on your system. Supported formats include:

• Word: .doc, .docx, .docm, .odt, .rtf, .pages
• PowerPoint: .ppt, .pptx, .pptm, .odp, .key
• Spreadsheets: .xls, .xlsx, .xlsm, .ods, .csv, .tsv, .numbers

Installation Commands:

• macOS: brew install --cask libreoffice
• Ubuntu/Debian: apt-get install libreoffice
• Windows: choco install libreoffice-fresh (Note: You may need to add the LibreOffice program directory to your PATH).

Images

LiteParse provides native support for image formats without requiring imagemagick (as of v2.8.0). Supported formats:

• .jpg, .jpeg, .png, .gif, .bmp, .tiff, .webp, .svg

For higher accuracy, you can replace the default Tesseract engine with an external HTTP OCR service (like EasyOCR or PaddleOCR).

To integrate your own service, implement the LiteParse OCR API specification. Your service must provide:

• A POST /ocr endpoint.
• Support for file and language parameters.
• A JSON response in the following format:

{
  "results": [
    {
      "text": "string",
      "bbox": [x1, y1, x2, y2],
      "confidence": number
    }
  ]
}

Complexity is computed per page. A page is flagged with a needs_ocr verdict if it contains any of the following reasons:

The JsParseResult is the top-level object returned after a parsing session.

Fields:

• total_pages: Total pages in the source document.
• pages: Array of JsParsedPage objects.
• page_errors: Array of JsPageError objects (if continue_on_page_error is enabled).
• text: The concatenated text of the entire document.
• images: Extracted image data.
• screenshots: Rendered page images (if extract_screenshots is enabled).
• doc_meta: Document-level metadata (if extract_document_metadata is enabled).
• xfa_packets: Raw XFA packets (if extract_xfa_packets is enabled).

The LiteParse skill is designed to keep an agent's context window small and reduce costs by following these patterns:

• Parse once, search many: Instead of re-parsing a document for every query, the skill parses the document to a temporary file once, and all subsequent searches are performed against that file.
• Minimize round-trips: The skill fetches matches and their surrounding context in single commands and batches independent lookups to avoid spending multiple turns on search terms.
• Bound outputs: Results are capped to prevent a single lookup from flooding the agent's context window.
• Escalate to ranked search: When keyword matching is insufficient, the skill uses ranked search instead of attempting multiple keyword variants one by one.