Introduction

The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” For developers who need to run agents locally and process documents, directly reading PDF papers often presents pain points: messy layouts, formulas and figures that are difficult to parse, and the inability to directly ask questions about specific local content within a conversation flow. The read-paper plugin aims to solve these problems. It converts PDFs into structured, selectable text and provides a paper reading panel with question-and-answer capabilities in DSH’s conversational interface.

What Is This

read-paper is a paper reading panel plugin designed for DeepSeek Harness. It is maintained by developer louwenbo580 and is licensed under the MIT License. It parses PDF papers into structured HTML documents (including two-column layouts, headings, tables, and table of contents), provides selectable text and image rendering in a sidebar, and supports instant question-and-answer on selected content as well as project workspace functionality.

Core Features

The plugin provides the following main capabilities:

  1. PDF to Structured Text
    It runs based on the host’s Poppler toolchain, rebuilds two-column layouts, extracts headings (with case normalization and heading splitting), identifies tables (including geometric column detection and wrapped-cell handling), processes superscripts and subscripts, and handles the table of contents.

  2. Figure and Formula Rendering
    It uses pdftoppm and pdfimages to extract images and display formulas, rendering them as images. At the same time, formula text is preserved as hidden HTML spans to ensure searchability.

  3. Right-Side Reading Panel
    It provides a third column in DSH’s details pane. When space permits, it displays as a full column; when space is insufficient, it automatically collapses, supporting a floating layout and a compact Tab mode.

  4. Ask Dialog
    Selecting any text in the reading panel triggers a floating dialog. After entering a question (such as “Explain this logic”), the selected text, a context summary, and the paper abstract are sent to the model for an answer.

  5. Projects
    Using the “Create project” feature, a paper (PDF, MD, HTML, images, README) can be materialized into a Paper workspace. This opens a new session and lets the agent load the paper as context, supporting multiple coexisting projects.

Installation and Enabling

Before installing, ensure DeepSeek Harness (dsh) is installed.

  1. Install the Plugin
    Use the following command to install the plugin into a specified profile (for example, readPaper):
    dsh plugin --profile readPaper add @deepseek-ai/dsh-web-app paper-review

*Note: `@deepseek-ai/dsh-web-app` includes the browser GUI layer required for plugin rendering.*

  1. Install Dependency Tools
    The plugin depends on Poppler command-line tools. Ensure that pdftotext, pdftoppm, and pdfimages are added to your system PATH.

    • macOS: brew install poppler
    • Debian/Ubuntu: sudo apt-get install poppler-utils

  2. Start Harness
    Start it using the configured profile:

    dsh --profile readPaper

Typical Usage

After installation and startup, you can use it through the following steps:

  1. Open the “Paper” tab on the right to enter the reading column.
  2. Click the “Load PDF…” button to upload a file, or drag and drop a PDF directly into the reading column.
  3. Select a passage in the converted text; an Ask dialog pops up. Enter a question and press Enter; the model answers based on that passage and its context.
  4. Click “Create project” in the toolbar; the system converts the paper materials into a Paper workspace and starts a session that automatically loads that paper as context.

Applicable Scenarios and Notes

  • Applicable Scenarios: Developers who need to deeply read arXiv or academic PDFs and want to use large language models to explain or summarize local content.
  • Note 1: Conversion fidelity heavily depends on the PDF’s text layer. For scanned PDFs, the plugin degenerates into rendering page images.
  • Note 2: Display formulas are rendered as images, and their mathematical text is hidden as spans, so the text in the images cannot be extracted via OCR.
  • Note 3: Content that exists only in PDF annotations (such as full author lists in some journals) cannot be extracted.
  • Note 4: The plugin’s HTTP routes follow DSH’s local trust model and require no additional configuration.

Conclusion

read-paper provides DSH users with an intermediate stage from simple conversation to in-depth document analysis. It reduces the barrier to processing PDFs through structured parsing and strengthens the agent’s contextual understanding of papers using the Ask Dialog and Projects features. The related code and documentation can be viewed in its GitHub repository.