ferstar / dsh-tool-ocr

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