DeepSeek Harness (DSH) adopts a plugin-based architecture and enhances model capabilities through extension tools. Vision models usually handle standard image formats well, but they often cannot directly read TIFF/TIF files, a format commonly found in scanned documents, faxes, and scientific data. The dsh-tool-read-tiff plugin is designed to fill this gap, expanding the model’s capability from “only looking at images” to “understanding scanned documents.”

Plugin Positioning

This is a read_tiff tool for models, maintained by developer fulander0301 and categorized under “Model Inference.” It can perform core decoding and conversion without relying on complex configuration, while providing complete metadata and supporting optional vision model description.

Core Features

The plugin provides the following capabilities:

  1. Input Support: Supports local absolute paths or HTTP(S) URLs as input sources.
  2. Decoding and Conversion: Decodes TIFF/TIF files and converts them into previewable PNG images. The converted path is returned via convertedPath.
  3. Metadata Extraction: Provides complete TIFF header metadata, including dimensions, page count, compression method, photometric interpretation, bit depth, and more.
  4. Multi-Page Processing: Supports multi-page TIFF files and allows access to specific pages through the page index parameter.
  5. Optional Visual Description: If an OpenAI-compatible vision endpoint is configured, the plugin sends the decoded PNG to the model to obtain a text description.
  6. Real-Time Configuration: Supports modifying endpoint configuration in real time via the settings card without restarting the process.

Installation and Activation

The installation process is completed through DSH’s plugin management command.

dsh plugin --profile web add git+https://github.com/fulander0301/dsh-tool-read-tiff.git

After installation, restart the dsh web process for the plugin to take effect. The plugin depends on the sharp and schemastery libraries, which each have their own licenses.

Typical Usage

The plugin can decode and convert by default without configuration. Enabling visual description requires additional configuration of an OpenAI-compatible endpoint.

Basic Usage Examples:
- “transcribe all text in this scanned fax”
- “extract the table as CSV”
- “describe the map features and projection”
- “what does the label in the corner say”

Multi-Page File Processing:
Since each call handles one page by default, multi-page files require multiple plugin calls with the page index specified (starting from 0). For example, first ask for pages to obtain the total number of pages, then call page: 0, page: 1, and so on.

When No Vision Endpoint Is Configured:
If no vision model endpoint is configured, the plugin still returns the decoded result and convertedPath. It typically returns a note prompting the model to use the installed describe_image tool to describe the image pointed to by convertedPath.

Notes

  1. Single-Page Limit: Each call handles only one page. Multi-page files must be processed with repeated calls.
  2. Security: The plugin rejects HTTP redirects. Raw TIFF and PNG bytes are not returned in the conversation; only text and metadata are returned, ensuring that sensitive image data is not directly exposed in logs.
  3. File Size and Format: Extremely large scientific raster TIFFs, or those containing complex floating-point multispectral bands, may require preprocessing; otherwise, the tool returns clear error messages.
  4. Input Restrictions: Only local absolute paths or http(s) URLs are accepted. Other URL protocols are rejected.

Summary

dsh-tool-read-tiff is a practical tool in the DeepSeek Harness ecosystem for handling TIFF document parsing. Through decoding and metadata extraction, it enables text models to process scanned documents, and its optional visual description feature supports image-text interaction. View more details or the source code in its GitHub repository or Skill Hub directory.