ferstar/dsh-tool-ocr

Name: ferstar/dsh-tool-ocr
Author: ferstar

插件Plugin 原生Native ⭐ 4 MIT 工具Tools 视觉Visual

本地 OCR 插件：让纯文本生成 LLM 也能读懂图片 | Local OCR plugin: give text-only generative LLMs the ability to read images

项目介绍Project Overview

由知微Agent(zhiwei)编辑Edited by Zhiwei Agent (zhiwei)

dsh-tool-ocr 是面向 DSH 模型的本地图片文字识别插件，通过独立的 nbocr 引擎（PP-OCRv6 模型）实现，无需视觉模型即可提取图像文字。支持 recognize / status / install / check 操作，输出按阅读顺序排列的文本、文本块坐标及低置信度/数字/日期等审阅标记，并可生成启发式 Markdown 表格。适用于无视觉通道场景下截屏、票据或文档图像的本地 OCR。注意：必须先安装 nbocr 可执行文件并通过 cordis.patch.yml 注入 command 路径。

dsh-tool-ocr is a local image OCR plugin for DSH models that lack vision input, powered by the standalone nbocr engine (PP-OCRv6). It exposes recognize, status, install, and check actions, accepting a local file path or an in-session attachment_id. Output includes reading-ordered text, per-block bounding boxes, review flags (low confidence, amounts, numbers, dates, quantities), and heuristic Markdown tables. The plugin registers a tool-ocr settings card on dsh ≥ 0.1.0-rc.7, letting you tune engine command, model tier, and run limits live without editing the composition. Note: you must install nbocr separately and supply its path via command in cordis.patch.yml.

在 GitHub 查看View on GitHub

发给你的 DSH 安装Send to your DSH

请帮我了解并安装插件：【dsh-tool-ocr】【https://github.com/ferstar/dsh-tool-ocr】

把上面这条消息直接发给当前会话里的 DSH，让它帮你了解并安装。安装命令不一定准确，发给 DSH 更稳。Send this message to DSH in your current session. CLI install commands may not be accurate across systems — DSH will figure it out for you.

或使用命令行安装（适合开发者）Or use CLI install (for developers)

命令行安装CLI Install

dsh plugin --profile web add dsh-tool-ocr

把 ferstar/dsh-tool-ocr 加入你的 DSH 配置（web profile）即可启用。

READMEREADME

3.7 KB · 缓存于 2026-08-23 03:42:40 · 3.7 KB

dsh-tool-ocr

English | 中文

Local image text recognition for DeepSeek Harness models without vision input, backed by the standalone newbee-ocr (nbocr) engine over PP-OCRv6 models. Fully out-of-tree: depends only on published dsh base packages.

Features

Item	Description
`ocr` tool	`recognize` / `status` / `install` / `check` actions
Image inputs	Local `path`, or `attachment_id` (an image already in the conversation)
Output	Reading-ordered text, per-block bounding boxes, review flags (low confidence / amounts / numbers / dates / quantities), heuristic Markdown tables
Robustness	Survives MNN diagnostics in engine stdout; honors cancellation and timeout
Settings card	On dsh ≥ 0.1.0-rc.7 the plugin registers the `tool-ocr` namespace: the web Settings → Plugins page renders a card that edits the engine command, model tier, and run bounds live — no composition edit needed

Quick start

1. Install the engine — grab nbocr from the newbee-ocr-cli releases (one-line installer or prebuilt archive):

curl -LsSf https://github.com/zibo-chen/newbee-ocr-cli/releases/latest/download/newbee_ocr_cli-installer.sh | sh
# Windows PowerShell: irm https://github.com/zibo-chen/newbee-ocr-cli/releases/latest/download/newbee_ocr_cli-installer.ps1 | iex

2. Install the plugin

dsh plugin --profile web add dsh-tool-ocr
# or: cd ~/.dsh/profiles/web && pnpm add dsh-tool-ocr

3. Mount it in your profile's cordis.patch.yml:

- insert:
    - id: ocr
      name: 'dsh-tool-ocr'
      inject: [tools, subprocess, systemPrompt]
      config:
        command: 'C:/path/to/nbocr.exe'   # required — the nbocr executable

That's it — restart dsh, then point the model at an image: ocr { path: "C:/screenshot.png" }. To tune recognition (language, model tier, limits), open Settings → Plugins → Plugin config and edit the OCR card; changes apply on save, no restart needed. command is the only field that must come from the composition.

Image inputs

path works on every dsh build: the model reads the file from disk.
attachment_id needs a build that admits images for text-only models (e.g. the dsh fork with api-gateway.allowImagePlaceholder: true): images enter the session, the llm layer substitutes [image attachment <id>] text, and the model hands the id to ocr.

Configuration

Field	Default	Meaning
`command`	— (required)	The nbocr executable: absolute path or PATH-resolved name.
`args`	`[]`	Extra arguments before the nbocr subcommand (no shell).
`env`	`{}`	Extra environment entries.
`language`	`chinese`	Recognition model/language alias.
`detModel`	`v6-tiny`	Detection model tier.
`modelsDir`	`''`	Model directory for non-embedded models; empty uses embedded models.
`maxImageBytes`	`26214400`	Largest accepted image in bytes.
`maxOutputBytes`	`2000000`	Largest collected engine stdout in bytes.
`maxTextChars`	`12000`	Largest recognized text returned in `text`.
`timeoutMs`	`600000`	Tool-call timeout budget in ms.

Tool usage

ocr { path | attachment_id, action?, include_boxes?, table?, max_text_chars? }

status probes readiness without engine work; check runs a real end-to-end 1x1 probe.

Development

pnpm typecheck   # tsc --noEmit
pnpm test        # vitest
pnpm lint        # oxlint

Real-engine smoke (skipped unless both env vars are set):

NBOCR_BIN=/path/to/nbocr OCR_E2E_IMAGE=/path/to/image.png pnpm test

License

MIT

渲染基于仓库 ferstar/dsh-tool-ocr 的 README.md。在 GitHub 查看原文 →

命令行安装CLI Install 复制安装命令Copy Install