dsh-doc-reader

Read PDF, DOCX, XLSX, PPTX and text documents as agent tools for DeepSeek Harness (dsh).

The built-in read tool returns file text, and web_fetch rejects application/pdf — so a PDF, a spreadsheet or a slide deck currently ends up as curl + pip install pypdf inside the session. This plugin closes that gap with five typed tools that return page-, sheet- and slide-aware text, including local OCR for scans and screenshots.

Screenshots

doc_read — page-aware PDF text and spreadsheet rows:

doc_ocr — local OCR for scans and images, with per-page confidence:

Tools

Every tool accepts a local path or an http(s) URL. Prefer doc_meta first for large files, then read the pages/sheets you need — output is capped by maxChars (default 60000, hard cap 400000).

Example output

# invoice.pdf — PDF · 184.2 KB
# source: /work/invoice.pdf
# pages 1, 2 of 7 · title: Invoice 2026-10

--- page 1/7 ---
ACME Ltd — Invoice 0042
...

# budget.xlsx — XLSX · 12.4 KB
# source: https://example.com/budget.xlsx
# sheets: Data, Second

## sheet "Data" — 3 rows × 3 cols
1 | Year | Revenue | Note
2 | 2024 | 1000    | ok
3 | 2025 | 2000    | Привет

OCR

doc_ocr handles what pure text extraction cannot: scanned PDFs, photos and screenshots.

• Engine: tesseract.js 7 (WebAssembly, no native OCR binary) — runs fully locally, nothing is uploaded.
• PDF pages are rasterized with pdf-to-img (pdf.js + @napi-rs/canvas) at scale 2 (~200 DPI), then recognized with one shared worker.
• Fax-style CCITT G4 and JBIG2 scans are decoded through pdf.js wasm/JS decoders (wasmUrl is wired to the installed pdfjs-dist/wasm), so two-tone scans rasterize instead of returning blank pages.
• Default languages: rus+eng (langs overrides, e.g. "deu", "eng"). Order matters: with eng+rus tesseract reads uppercase Cyrillic as Latin lookalikes (ПРИВЕТ → NMPUBET); Cyrillic-first keeps Latin accuracy identical (measured 94–95% on both fixture sets).
• First use downloads the language data (~4 MB per language) into ~/.dsh/cache/dsh-doc-reader/tesseract; later calls are fast (a page takes ~0.3–1 s). For offline machines set DSH_DOC_READER_OCR_LANG_PATH to a directory containing .traineddata (optionally .gz), and/or DSH_DOC_READER_OCR_CACHE for the cache location.
• At most 20 PDF pages per call; narrow with pages. Use scale: 1 for speed, scale: 3 for small print.
• Rotation/handwriting/low-resolution scans are best-effort: tesseract has no layout magic, and confidence is reported per page so you can judge the result.

# scan.pdf — PDF (OCR) · 1.8 MB
# source: /work/scan.pdf
# OCR: tesseract rus+eng · pages 1, 2 of 12

--- page 1/12 (ocr, 91.4%) ---
INVOICE 2026-042
...

Limits and security

• Local files: up to 50 MB; URLs: downloaded with a 20 s timeout and the same 50 MB cap.
• Private/local hosts are refused (localhost, 127.0.0.1, RFC1918, .local, .internal) so the tool cannot become an SSRF primitive. Download the file and pass a path instead.
• Only http(s) URLs.
• PDF text extraction is text-based: scanned PDFs (images only) return (no extractable text on this page) — use doc_ocr for those.
• DOCX conversion is handled by mammoth; charts, comments and tracked changes are not materialized as text.
• XLSX returns displayed values, not formulas; rows are paged with maxRows.

Install

Published on npm as dsh-doc-reader. Install from the built-in plugin market, or manually into a profile:

cd ~/.dsh/profiles/desktop
npx pnpm add dsh-doc-reader

Then add the bundle to the profile's package.json:

"dsh": { "profile": { "bundles": ["dsh-doc-reader", "..."] } }

Restart DSH (the plugin loads on boot; DSH composition is all-or-nothing). The five tools appear in the next session.

For local development, install the working copy instead: npx pnpm add link:~/dsh-plugins/dsh-doc-reader.

> pnpm 12 may report an ignored build script for tesseract.js. That postinstall is only the OpenCollective banner — approve it with pnpm approve-builds or ignore it; OCR works either way.

Develop

npm install
npm test          # node --test, runs against test/fixtures
node --check index.js

The readers live in lib/ and have no host dependency, so they can be exercised directly:

import { readDocument, documentMeta } from "./lib/read.js";
await readDocument({ path: "report.pdf", pages: "1-3" });

index.js imports @deepseek-ai/dsh-tools for defineTool; outside a DSH host the import falls back to an identity definition (the registry consumes the plain object).

Roadmap

• .doc, .rtf, .odt, .epub readers.
• Layout hints from OCR (word boxes, tables) and page-number citations.
• Page/sheet images for vision models (presentResult image parts).
• Caching of parsed documents within a session.

License

MIT