Introduction

The core idea of DeepSeek Harness (DSH) is “everything is a plugin.” When working with DSH, reading PDF documents usually requires manually copying text or taking screenshots before AI analysis, which disconnects the reading and conversation flow. The dsh-pdf-reader plugin aims to integrate a PDF reader, annotation tools, and AI analysis in the same web workflow.

What It Is

dsh-pdf-reader is an external (out-of-tree) plugin maintained by user ydlstartx. It provides AI-driven PDF reading capabilities for DeepSeek Harness, supporting native text, image regions, visual assets, multi-PDF joint questioning, and on-demand OCR after user authorization.

Core Features

The plugin mainly includes the following capabilities:

  1. Reading and navigation: Supports multiple PDF tabs, table of contents, bookmarks, thumbnails, full-text search, jump history, zoom, fit, and lossless rotation.
  2. Text and annotation: Provides native text selection, copying, and AI operations; supports highlighting, underlining, strikethrough, notes, and annotation center management.
  3. Image and text AI: Ask questions about image regions, add charts, captions, and related paragraphs to a shared visual asset, perform joint analysis, and trace back to the original source through source links.
  4. Whole-document Q&A: Supports single-PDF or multi-PDF joint questions, fixes the current document scope for the session, and distinguishes facts from inferences by source in the answer.
  5. On-demand OCR: Disabled by default; after explicit user authorization, the AI recognizes only a small number of pages as needed. The OCR page cache is limited (about 64 MiB or 512 pages).
  6. Save and restore: Restores opened documents and reading state per conversation; supports “Save As” and “Overwrite Original After Confirmation.”

Installation and Enablement

The plugin is released as a Beta version (current version 0.1.0-beta.3) and requires DeepSeek Harness 0.1.1-rc.2.

After downloading the prebuilt .tgz package, run the following commands in the root directory of the Harness checkout:

pnpm dsh plugin --profile web add /absolute/path/to/dsh-pdf-reader-0.1.0-beta.3.tgz
pnpm dsh --profile web --dump-config | rg dsh-pdf-reader
pnpm dsh web

Typical Usage

  1. Ask about an entire PDF: Select one or more PDFs and ask directly. OCR is disabled by default and only recognizes on demand after authorization.
  2. Select text and invoke AI: After selecting text, you can copy, explain, translate, or add it to the context.
  3. Create and save annotations: Supports highlights, underlines, notes, and more; annotations are written to the PDF only when “Save As” or “Overwrite Original After Confirmation” is performed.
  4. Combine images with original text to ask: Add charts to the visual asset basket, analyze them together, and get clickable source backlinks.
  5. Sample assets and authorization: When image regions need to be analyzed with a visual model, ensure the current model supports visual input.

Use Cases and Notes

  • Use cases: PDF users who need to read and analyze within a conversation, especially when dealing with charts, multi-document comparison, or citing original text.
  • Runtime requirements: A modern Chromium browser that supports IndexedDB, Web Worker, Canvas, and Blob URL is required.
  • Permissions and privacy: The plugin does not modify or require customization of the DSH source code. Data (PDFs and annotations) is stored in the browser’s local storage for the current site. It does not upload or modify PDFs unless the user explicitly invokes an AI feature or confirms overwriting a file.
  • Security constraints: PDF content, annotation text, and images are treated as untrusted data and are not executed as system instructions.
  • Known limitations:
    • Selection boxes for image regions cannot span pages.
    • The table of contents and sidebar search only read information embedded in the PDF and do not trigger OCR.
    • The first OCR run requires a network download of the PaddleOCR model and will fail offline without a cache.
    • The 300 MiB limit is a defensive cap at the file entry point and does not represent a committed large-file performance metric.
    • A visual model is required to analyze image regions.

Summary

dsh-pdf-reader tightly integrates a PDF reader with AI analysis and solves evidence-chain tracking with a visual asset basket and source backlinks. For developers who need to process PDF documents in DeepSeek Harness, this is a practical, ready-to-use tool.