Introduction

DeepSeek Harness (DSH) uses a plugin architecture, allowing community developers to extend functionality. rangrongg/searchsieve belongs to the web search tool category and is designed to address the challenge of quantifying search result quality in RAG (Retrieval-Augmented Generation) scenarios. It filters and evaluates search results through semantic analysis, helping developers intuitively determine whether retrieval is sufficient.

Core Features

The plugin provides three core metrics: retention ratio, answer completeness, and representativeness and redundancy. These metrics clearly convey search result quality in one go.

Each time a web search is initiated, the coverage card appears automatically, without manually assembling parameters. The card supports switching among three skins: pink-white, blue-white, and pure white. Click the top-right corner of the expanded card to switch; the selection is saved automatically.

This tool uses a pure rule-based algorithm, does not call models, and does not consume API keys. It features zero model overhead and deterministic output. It is adapted for Chinese-language environments and can verify semantic points one by one without a word-segmentation dictionary. In addition, the plugin handles edge cases well: empty results, zero selections, and other situations all include clear prompts and suggestions.

Installation and Activation

The installation command is as follows:

dsh plugin add search-sieve

After installation, restart dsh web. After the restart, the coverage card will appear automatically the next time a web search is initiated.

Usage

The plugin accepts an expectedPoints (expected answer points) parameter. When this parameter is provided, the tool verifies and evaluates the results point by point. If it is not provided, the tool automatically extracts keywords from the question.

A sample usage is as follows:

{
  "query": "村上春树一共写了几本书",
  "expectedPoints": ["作品数量", "最新作品", "诺贝尔奖经历", "代表作书名"],
  "searchResults": [
    "村上春树共出版 14 部长篇小说，最新作品《城市及其不确定的墙》 https://example.com/a",
    "村上春树多次入围诺贝尔文学奖但未获奖 https://example.com/b"
  ],
  "selectedResults": [
    "村上春树共出版 14 部长篇小说，最新作品《城市及其不确定的墙》 https://example.com/a"
  ]
}

The plugin returns JSON data containing the following fields:
- selectionRatio: selection ratio
- answerCompleteness: answer completeness
- representativeness: representativeness and redundancy relative to the full result set
- gaps: uncovered information points
- suggestion: supplemental search suggestion

Notes

  1. Version requirement: dsh web 0.1.0-rc.x or a newer version is required.
  2. Algorithm principle: pure rule-based algorithm; no model calls; no API key required.
  3. Chinese support: no word-segmentation dictionary required for Chinese.
  4. Windows environment: If Windows PowerShell reports that script execution is disabled, use dsh.cmd or npm.cmd instead.