Introduction

DeepSeek Harness offers a plugin-based extension capability. When handling document tasks, common approaches are pasting the full document into the prompt (limited by the context window) or invoking it through external scripts. The dsh-evidence plugin addresses this: it allows you to upload files in the Harness Web composer, with parsing and indexing performed on the local machine, while the model retrieves compact evidence through tools and expands versioned, precise coordinates when needed.

Plugin Overview

dsh-evidence is a community plugin for DeepSeek Harness, maintained by Cooberped.

Core value: It treats files as retrievable evidence instead of stuffing large blocks of text into the context. Parsing and indexing are performed locally, and raster images are processed through Harness’s native vision path.

Features

  • Local parsing and indexing: File parsing and indexing are completed on the local machine, and full documents are not sent to the cloud.
  • Versioned coordinates: Retrieved coordinates include version information (source hash + parser/block architecture). If the source file changes, old coordinates become invalid.
  • Compact evidence retrieval: The model calls tools to find evidence instead of copying the entire document.
  • Multi-format support: Supports PDF, DOCX, XLSX, PPTX, and native raster images.
  • Two core tools:
    • search_documents: Finds the “location” where the answer is.
    • read_document: Reads the “content” at specific coordinates.

Installation and Enabling

Before installing, ensure your environment meets the following requirements:
* DeepSeek Harness version >= 0.1.7
* Node.js version >= 22.13.0

Install using npm:

dsh plugin --profile web add @baseland/dsh-evidence@beta

After installation, it is recommended to run the following commands to verify that the configuration has taken effect and restart Harness:

dsh --profile web --dump-config
dsh web

Typical Usage

  1. Open the Harness Web composer.
  2. Add files: use the paperclip to select multiple files, use the folder button to upload a directory, or drag and drop files directly onto the page.
  3. Confirm that each file has its own card displaying “AI-readable”.
  4. Ask a question. The model will automatically call tools to retrieve information.

The following are effective prompt examples:

Index these three files first. Do not summarize them yet.

Across these files, find the definition and formula for the Q3 retention metric.
Give the source file and exact page, slide, line range, or Sheet!Range for every claim.

Compare the meeting decision with the workbook target.
If the documents do not contain enough evidence, say what is missing instead of guessing.

Notes

  • Version tag: The plugin is currently in the public npm beta phase. When installing, it is recommended to explicitly specify the @beta tag.
  • Coordinate versioning: Coordinates include version information. If the source file is modified, old coordinates cannot be read correctly, and the model will receive an error message.
  • Zero-recall rule: When no evidence can be found in the files, zero recall does not mean the model is allowed to guess. The model should report missing evidence or suggest using different search terms.

Summary

The dsh-evidence plugin uses local parsing and versioned coordinates to address context management and accuracy issues when the model reads files. It transforms files into structured, traceable evidence, making it suitable for document analysis tasks that require precise source citations.