@zhtx2026/dsh-pdf
PDF parsing toolkit for DeepSeek Harness — three agent-facing tools (pdf_info / pdf_extract_text / pdf_render_page) with hybrid engines, including system-font rendering for PDFs with non-embedded CJK fonts (e.g. EasyEDA / JLC EDA schematic exports).
English | 中文
Tags: dsh-plugin · pdf · pdfjs · canvas · schematic · tools
Install
dsh plugin --profile web add @zhtx2026/dsh-pdf
Or from a local checkout during development:
dsh plugin --profile web add link:<path-to-repo>
Then restart DSH (or open a new session) — the three tools and an agent guidance block become available.
Tools
| Tool | What it does |
|---|
pdf_info | Page count, page sizes (pt), document metadata, and the font list with per-font embedded flags. |
pdf_extract_text | Extracts text per page; optional page and withPosition (per-line x/y coordinates + font size). |
pdf_render_page | Renders one page to a PNG (local file) with configurable scale/outDir; engine auto-selection. |
Why a custom renderer?
pdfjs-dist cannot render text in PDFs whose fonts are not embedded and lack a
ToUnicode map — the classic case is JLC EDA / EasyEDA schematic exports
(SimSun/SimHei Type0 fonts with UniGB-UCS2-H encoding): pdfjs emits
translateFont failed and drops the glyphs.
dsh-pdf solves this with a two-engine design:
- pdfjs engine — used when all fonts are embedded (standard PDFs).
- sysfont engine — a built-in content-stream renderer that decodes the
PDF text operators (
BT/ET, Tf, Tm, Td, Tj, TJ) itself, maps
UniGB-UCS2-H/UTF-16BE strings to Unicode, and draws them with system
fonts (SimSun, SimHei, Microsoft YaHei, …) via @napi-rs/canvas.
Text extraction uses the same auto-selection: embedded fonts → pdfjs
text layer; non-embedded fonts → the built-in decoder (which correctly
recovers CJK text pdfjs loses).
Example
pdf_info("D:/Downloads/SCH_SA-V11A.pdf")
→ 3 pages, A3 landscape, jsPDF producer, 6 fonts (all non-embedded)
pdf_extract_text("D:/Downloads/SCH_SA-V11A.pdf", { page: 1 })
→ net names, pin numbers, and Chinese labels ("主控板", "嘉立创", …)
pdf_render_page("D:/Downloads/SCH_SA-V11A.pdf", { page: 1, scale: 2 })
→ D:/Downloads/SCH_SA-V11A-pages/page-1.png (2396×1698, engine: sysfont)
Architecture
lib/index.js — cordis plugin entry (name, inject, apply); registers
the three tools as native tool objects (plain JSON-Schema parameters and
output — no @deepseek-ai/dsh-tools dependency needed) and injects the
agent guidance block (systemPrompt.section, order 160).lib/pdf-core.js — self-contained engine: PDF object parser, font scanner,
pdfjs loading with fs-based CMapReaderFactory/StandardFontDataFactory
(Node 24's global fetch doesn't speak file://), the built-in text
extractor, and the sysfont renderer.cordis.patch.yml — the dsh.bundle.patch insert row.
Limitations
- Scanned/image-only PDFs have no text layer — extraction returns nothing
(rendering still works, as a plain image).
- The sysfont renderer assumes
Type0 CID codes equal Unicode code points
(true for UniGB-UCS2-H exports); exotic CID fonts fall back to pdfjs.
- Large PDFs are parsed fully into memory.
License
MIT
