A drawing PDF is not a document — it is text, vector geometry and dimension strings scattered across a sheet. PDFWhisperer reads all three with their positions intact, so a callout stays tied to the thing it points at and a dimension arrives as a value with units, not loose characters for something downstream to re-guess.

Why not just copy-paste the PDF text: Selecting text off a sheet gives you a flat wall of characters with the layout thrown away — and gives you nothing at all when the page is a scan. PDFWhisperer keeps each item's position and font size, parses dimension strings into values, and rasterizes-and-OCRs the pages that carry no decodable text layer, so a scanned sheet takes the same path as a native PDF instead of failing quietly.

What it extracts today

◈

Text with position

Every text item keeps its coordinates and font size, so titles, callouts and schedule cells stay distinguishable from body notes rather than collapsing into one stream. (POST /extract)

◈

Geometry

Vector content on the page is read alongside the text — which is what lets a callout be tied to the thing it points at, rather than floating free.

◈

Parsed dimensions

Dimension strings are parsed into values and units instead of being left as loose text for something downstream to re-guess.

◈

OCR fallback for scans

Some PDFs use fonts with no ToUnicode map and cannot be decoded at all. Those pages are rasterized and read by OCR, so a scanned sheet is handled by the same path as a native PDF. (POST /ocr — OCR runs on the RedKitten network, not on your machine.)

◈

Table extraction

Schedules and title-block tables are pulled out as structured rows and columns rather than a run of cells. (POST /extract-table)

◈

Ask the page a question

The extracted bundle can be put to a language model together with your own notes, so a question about a sheet is answered against that sheet's real contents. Inference runs on the RedKitten DGX gateway (remote compute), not locally. (POST /ask)

How the output travels

A structured bundleThe primary output is plain structured data (JSON). Anything that can read JSON can consume it — no proprietary format.

A model promptThe same bundle renders into a prompt so a language model reasons over the sheet's real contents instead of a guess at them.

Position-aware, not flatBecause every item carries its place on the sheet, the extraction preserves the layout a plain text copy throws away.

Reading aid, verify against the sheetExtraction and OCR are aids. Anything you rely on should be checked against the drawing.

Submittal reviewComparing contractor drawings against the contract documents and surfacing deviations, omissions and mismatches as markups on the sheet. This is the product direction; the compare-and-review workflow is not built yet.

Hosted web interfaceA browser UI on top of the extraction service. Sign-in is already wired (Google via SiteGate); what's missing is the interface behind it. Today the service is reached as a backend on the RedKitten network.

Desktop appA Tauri desktop build exists in-repo but is not distributed yet.

DWG inputA bridge for reading DWG files through the same extraction path is designed but not yet shipped.

Availability

PDFWhisperer runs today as a backend service on the RedKitten network. A hosted, signed-in way to use it — and the submittal-review workflow — is in development.

Hear when access opens

PDFWhisperer runs today as a backend service. Leave your email and you'll hear when a hosted, signed-in way to use it — and the submittal-review workflow — is ready.