Why Plain Text Extraction Fails on a Drawing

A scour-depth requirement is sitting in the corner of sheet C-14, hand-noted by the structural engineer during a late design revision. Three weeks later, an RFI comes in asking exactly what that note says. Someone runs it through your document search. Nothing comes back. Not because the note isn't there. It's right there, legible, on the sheet everyone has open. It's because whatever indexed that sheet never actually read it. It saw a picture. That's the gap this article is about.

Every drawing in your bid package is a PDF, and every PDF looks the same from the outside. But there's a difference between a PDF with real text underneath it and a PDF that's just a photograph of a page, and that difference is invisible until you try to search it.

A text layer is what makes a PDF's words findable: it's the copy-pasteable version sitting underneath the picture of the page. Most everyday extraction tools only ever read that layer. On an office memo or a spec document, that layer is usually there and usually complete, so the tool works.

Drawings break that assumption three different ways, and a typical bid package usually contains all three:

-A clean CAD export looks perfect, but its text layer is nearly empty. The pile marks, the zone colors, the schedule grid, the layout: all of that meaning lives in the drawing's visual structure, not in the extractable text, so a text reader hands back a handful of stray labels with no context.

-A plotted, image-only export has no text layer at all. Ask a text reader to find anything on it and it comes back empty, every time.

-A scanned legacy standard is the worst case: no text layer, plus handwritten annotations and hand-drawn markups that a plain extraction pipeline was never built to notice.

In all three cases, a person looking at the sheet sees everything. A text extractor sees nothing. That gap is where the note on sheet C-14 goes missing.

One more failure mode trips up tools that otherwise seem to work: color. On a lot of civil and structural sheets, color isn't styling. It's the encoding. A soil-zone map, a scour-depth diagram, a pile-quantity key: each one uses color to carry a specific engineering meaning, this shade means this condition. A tool that reads the labels but throws away which color they're attached to has thrown away the part of the drawing that actually says something. Axion treats that color-to-meaning mapping as content to be captured, not styling to be ignored.

Axion doesn't scrape a text layer. It looks at the drawing as an image, the way a person does, region by region. It renders each page at high resolution so small type and dense line work survive, then identifies the meaningful zones on the sheet (the title block, the legend, the schedules, the drawing field itself) and reads each one for what it specifically holds. The legend and title block get read first, so by the time the rest of the sheet is interpreted, Axion already knows what each color means and which sheet it's looking at.

Here's the part that matters most: the notes get captured too, not just the picture. Drawings carry their governing detail in writing, things like a design-basis wind speed, a corrosion allowance, a sequencing instruction, an exception that overrides the typical detail, and that text goes into the same search index and knowledge graph as everything else Axion reads. That corner note on sheet C-14 becomes as findable as a clause in a spec. Ask about it, and you get the actual requirement back, not a one-line description of the picture.

Schedules are usually the densest, most valuable content on a sheet (a foundation schedule with nested axes, a pile-quantity table, a cable schedule listing a size for every run) and they're also where a shortcut extraction does the most damage. Axion captures every column of a schedule verbatim: nothing dropped, nothing rounded, nothing inferred. Each row becomes its own structured, retrievable record tied back to its drawing, and Axion tags what each column means so you can filter by it. If it's genuinely unsure how to categorize a column, it still captures the column in full and flags the uncertainty. It never guesses and quietly drops what it couldn't place.

Full-fidelity capture is the setup. The payoff is what you can check it against. #

Once a schedule row or a drawing note is in the graph, Axion can compare it against your scope of work and your technical specifications. This is the same cross-reference it already runs on a contract body against its own exhibits to catch a mismatched liability cap or a domestic-content requirement buried three exhibits deep. Applied to drawings, that means the embedment depth on a foundation schedule gets checked against the geotech-driven requirement in the spec. A corrosion allowance noted on a structural sheet gets checked against the material class called out in the technical specifications. A pile- quantity key gets checked against the scope of work it's supposed to satisfy.

That's a materially different question than “can I search this drawing.” It's “does this drawing agree with the documents that govern it,” asked automatically, on every sheet, instead of relying on someone in the plan room to catch it by eye. The inconsistency that would otherwise surface as an RFI mid-construction, or as a change order once it's too expensive to fix cheaply, surfaces during bid review instead, while it's still a redline, not a delay.

Reading a drawing for its content and structure is not the same as certified quantity takeoff, and that distinction matters enough to say plainly. What Axion pulls from a drawing is structural and qualitative: it tells you what a sheet shows, what a note requires, what a schedule row contains. It is not a stamped source of precise counts or linear footages for procurement or estimating of record. Native CAD geometry (DWG, DXF) is also outside current scope; Axion reads drawings as delivered, not by ingesting CAD files directly. And on the hardest inputs, an illegible scan, a schedule with an unusually merged header, the column tagging can degrade, though the raw rows are still captured in full and flagged for review. The commitment is that the underlying data is preserved and its uncertainty is visible, not that every drawing yields a perfect structured read with no one checking it.

Reading each document well, drawings included, gets you a large pile of well-understood pieces. On its own, that's still just a better-organized pile. The value compounds when Axion connects those pieces, recognizing that the inverter on a schedule, the inverter in a submittal, and the inverter in a change order are the same inverter. Building that connected map is the subject of the next article.

Axion - Institutional Intelligence for Solar, BESS & Construction

Follow us on Axion for weekly technical insight.