dsh-pdf-reader

When vision models read PDFs they face a dilemma: whole-page rasterization buries small text, sub/superscripts, and vector-figure detail under a fixed pixel budget, while pure text extraction discards the spatial structure of figures, tables, and equations. dsh-pdf-reader resolves this with a per-page, content-aware strategy that feeds each page to the model in the mode that best preserves it.

Backed by PyMuPDF, the plugin exposes three tools. pdf_scan profiles every page (column count, vector regions, raster images, tables, formula risk, text-layer presence) so the agent can plan its reading order. pdf_read_page in mixed mode returns a low-res full-page preview, the extracted text, and auto-detected figure/table regions rendered as high-DPI PNGs stored in a local cache and referenced only by path, keeping heavy content out of the context window. pdf_render_region lets the agent zoom into any [x0,y0,x1,y1] rectangle at a DPI calibrated to fill the model's pixel budget without exceeding it. The result is a preview, content, refine loop that stays token-efficient and layout-faithful.

Built for developers whose AI agents need to closely read academic papers, technical whitepapers, or any PDF where figures, tables, and mathematical notation carry meaning. No manual per-page judgment or pixel math required: the plugin handles the routing automatically, and the agent only decides what to look at next.