OCR for construction documents does not work, we fixed it | Hacker News Reader

OCR for construction documents does not work, we fixed it

getanchorgrid.com

127 pointswcisco1798 comments

So we've built an API and trained models that detects fixtures, extracts schedules, and analyzes construction documents. Check us out!

More examples: - https://www.getanchorgrid.com/developer/docs/endpoints/drawi...

Main website: - https://www.getanchorgrid.com/developer

Why we did it: https://www.getanchorgrid.com/developer/docs/changelog/const...

Open articleView on HN

> OCR for construction documents does not work

I'm reminded of the Xerox JBIG2 bug back in ~2013, where certain scan settings could silently replace numbers inside documents, and bad construction-plans were one of the cases that led to it being discovered. [0]

It wasn't overt OCR per se, end-user users weren't intending to convert pixels to characters or vice-versa.

[0] https://www.youtube.com/watch?v=c0O6UXrOZJo&t=6m03s

If I recall it was an artifact of the compression algo.

Full context and details: https://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres...

JBIG2 does glyph binning, as you say not exactly OCR, but similar. So chunks of the image that look sufficiently similar get replaced with a reference to a single instance.

> not exactly OCR, but similar. So chunks of the image that look sufficiently similar get replaced with a reference to a single instance.

How can we describe OCR that wouldn't match this definition exactly?

It's not too hard, while they share some mechanics, the underlying use-cases and requirements are very different.

_______ Optical character recognition:

1. You have a set of predefined patterns of interest which are well-known.

2. You're trying your best to find all occurrences of those patterns. If a letter appears only once, you still need to detect it.

3. You don't care much about visual similarity within a category. The letter "B" written in extremely different fonts is the same letter.

4. You care strongly about the boundaries between categories. For example, "B+" must resolve to two known characters in sequence.

5. You want to keep details of exactly where something was found, or at the least in what order they were found. You're creating a layer of new details, which may be added to the artifact.

_______ "Glyph compression":

1. You don't have a predefined set of patterns, the algorithm is probably trying to dynamically guess at patterns which are sufficiently similar and frequent.

2. Your aren't trying to find all occurrences, only sufficiently similar and common ones, to maximize compression. If a letter appears only once, it can be ignored.

3. You do care strongly about visual similarity within a category, you don't want to mix-n-match fonts.

4. You don't care about clear category lines, if "B+" becomes its own glyph, that's no problem.

5. You're discarding detail from the artifact, to make it smaller.

Glyph binning looks for any chunks in the image that are similar to eachother, regardless of what they are. Letters, eyeballs, pennies, triangles, etc without caring what it is. OCR looks specifically to try and identify characters (i.e. it starts with a knowledge of an alphabet, then looks for things in the image that look like those.

If the image is actually text, both of them can end up finding things. Binning will identify "these things look almost the same", while OCR will identify "these look like the letter M"

Jbig2 dynamically pulls reference chunks out of the image, which makes it more likely to have insufficient separation between the target shapes.

It also gives a false sense of security when it displays dirty pixels that still clearly show a specific digit, since you think you're basically looking at the original.

That's a description of Jbig2, not a description of OCR.

Jbig2 is an OCR algorithm that doesn't assume the document comes from a pre-existing alphabet.

You asked what the difference was, and I said the difference. Was it unclear that to fit the phrasing of your question, we add "OCR doesn't"? I would not personally call Jbig2 OCR.

> You asked what the difference was, and I said the difference.

Take another look at my comment.

Let me try rephrasing to make the response to your original comment as clear as possible.

Question: "How can we describe OCR that wouldn't match this definition exactly?"

Answer: This definition largely fits OCR, but "reference to a single instance" is a weird way to phrase it. A better definition of OCR would include how it uses builtin knowledge of glyphs and text structure, unlike JBIG2 which looks for examples dynamically. And that difference in technique gives you a significant difference in the end results.

Is that better?

The definition you quoted is not an "exact" fit to OCR, it's a mildly misleading fit to OCR, and clearing up the misleading part makes it no longer fit both.

I cannot wait for the day when tech companies become players in the construction industry because it looks like it is the only way forward to make a change.

To think that everything has been digitalized a long time ago, yet contract law cannot properly deal with delineating responsibilities between GC and Architects, who are still sending 2D drawings to each other.

Imagine, all this information about quantities and door types (and everything else) is already available and produced by the architect's team, BUT they cannot share it! Because if they do, they are responsible for the numbers in case something is wrong.

So now there is this circus of: Arch technologist making the base drawing with doors. GC receives documents, counts doors for verification, and sends them to the sub. Subcontractor looks at these drawings, counts them again, and sends data to the supplier. Guess what, the supplier also looks, counts, confirms, and back we go.

Though I think robotics will change all of that. And when we have some sort of bot assistance, big tech players will have a bigger leverage in this, which will lead to the proper change management architecture.

Anyway, cool product. Anything to help with estimation. Really hope it gets traction.

I had a job as HVAC engineer for the upgraded Oslo Airport back in 2011; started doing HVAC work for 3 weeks the rest was programming trying to make the rest of people more efficient. Made an Excel sheet with a lot of macros to manage all the drawing of the airport. That’s why I switched to programming when I continued to study, and did not want to come back before I got more experience.

They even gave me a big desk at Trondheim/Tyholt so I could help them with the software during my studies.

For what it's worth I found Oslo Airport to be one of the best airport experiences I've ever had. 5 stars.

Thank you, I remember a lot of good discussions between architects and engineers. How to make it both beautiful and a good environment to be in

Hey thanks!! For estimation we cover division 08 (doors and opening) you can use it for your estimating purposes with these two endpoints:

- Counting all the doors: https://www.getanchorgrid.com/developer/docs/endpoints/drawi... - Extracting schedules in architectural drawings: https://www.getanchorgrid.com/developer/docs/endpoints/drawi...

and use Claude or any other AI tool to wire up the UI

We're releasing toilets (division 10) later this week, then floors and pipes next.

I’ve worked on projects where a lot of work was done in highly collaborative drawings on Bluebeam, in which vendors add their markups and items and the program facilitates counting it all at the end of the phase. My role was only in things like wireless AP placement and low voltage cabling drop locations, not anything safety critical like doors, but I assume those vendors were able to keep track of those items in a similar way. For actual engineering projects I’m glad so many people have to take the time to count.

I couldn’t agree more. The fact this data isn’t programmable is really holding the industry behind.

When building PlanGrid there were so many things we wished we could have done had this been unlocked.

I’m now working on doing just that.

We’re taking a different path, building a parsing engine that converts CAD (DWG/DXF) into fully structured JSON with preserved semantics (no ML in the critical path).We also have a separate GIS parser that extracts vector data (features, layers, geometries) independently, Like to know how you handle consistency and reproducibility across runs using models and how you make it affordable, especially at scale. because as far as i know CAD and GIS need precision and accuracy.

interesting yeah parsing DWG/DXF natively makes sense when the source file is clean and well-structured. The precision argument is valid in controlled environments.

The challenge we kept running into is that construction drawings in the wild aren’t always that clean. Unresolved xrefs, exploded dynamic blocks, version incompatibilities, SHX font substitutions — by the time a PDF hits a GC’s desk it’s often the only reliable artifact left. The CAD source may not even be available.

That’s why we see vision becomes the more pragmatic path — not because it’s more precise than structured CAD parsing, but because PDFs are the actual lingua franca of construction. Every firm, every trade, every discipline hands off PDFs. So we made a bet on meeting the document where it actually lives.

On consistency and reproducibility — that’s a real challenge with vision models. Our approach is to keep detection scope narrow and validate confidence scores on every output rather than trying to generalize broadly. Happy to go deeper on that if useful.

As a part of our product development, we had fought with PDF so much, even we have a generic PDF parser with triple pipeline (One for single column, another for multi column and third for complex table based layouts) yet we are not getting 100% accuracy, I would say that it's bit risky to bet on PDF. PDF often is the most complex format ever made and it was never made for data extraction. And You are right that vision models are the only way but hallucination is real.

Dumbcad line barf will not help you with that at all.

There already is a format that is plain text and preserves the semantics: IFC. That's what it was made for.

We’re not just dumping primitives, we extract full CAD context including entities, layers, blocks, colors, and topology. That metadata allow reconstruct structure deterministically. IFC is great when available, but in most real-world pipelines DWG is still the source of truth, often degraded. Our focus is making that usable without relying on probabilistic vision layers. People depend on PDF for cad files due to its portability and to avoid software dependency/licensing, we aim solving that, any machine or pipeline that needs CAD or GIS data for analytics, search, or reasoning can operate on our structured output without requiring a native CAD or ESRI license.

Is this a service / product you plan to offer outwardly? I'd be interested in learning more. Use case: estimation.

Happy to say that yes. We are in final round of polishing. mostly opening in couple of weeks. We are mainly targeting such uses cases where you can add CAD files into RAG or analytics or search pipeline without losing source of truth or geometry. I will definitely post here when we are ready, keep an eye and it will be free during beta, so you can play how much you want with it.

I think given the scale of projects worked on / being estimated (8 to 9 figure projects) -- it's unlikely we'd be willing to test on our real CAD files. Any chance for enterprise hosted solutions that aren't on the cloud?

That is the catch, yes on premise air gapped will be provided, this is purely inbuilt parsers no cloud dependency. We will also have a managed dedicated environment, if you don't want to manage the complex infra yourself. Even in Hosted model, we have provision to use your own s3 storage (only used for async parsing pipeline and file will be wiped after 3 hours)so file never touches our disc. We had considered this from a customer point of view and while designing the system. In my early career i had worked in GIS industry and know the privacy and the data security that is required for CAD and GIS files.

First off, congrats on the launch! Construction is a tough market to build in. My personal view after being in it a for a few years is that there is no shortage of MVPs. In fact there is an MVP for every problem at every level (or at least it feels that way) but construction is /vast/ and the rough edges that seem juicy at first, in practice are optimizations rather than bottlenecks for constructors.

I hope you succeed because it would be great to have a standard API for this data, but I would advise on one of two directions: become the standard by being close to 100% accurate at finding symbols (one symbol doesn't seem to cut it in our testing) or make a great, comprehensive workflow for a small subset of the market and become standard that way.

In both cases, you cannot do a broad 'market test', you need to spend many hours with a specific sub-set of users in construction.

Disclaimer: I'm a co-founder of Provision.

I ran the example doors given and it missed 9 swinging doors, some that were in double swing pairs, and a few that were just out on their own not clustered. Not bad overall though

Yep we're constantly improving we're currently above 0.87 for doors

we're thinking of adding a params for the ROC curve so that you can decide your own optimal thresholds depend on when false positive true positive rate is acceptable

Oh nice! I spend a good amount of time eyeballing drawings for overlooked details, so even finding most is a handy tool to me as my brain can skip the marked areas

Lovely. Are you in the construction space or are you a dev or both?

Construction, coding is a hobby from before I realized a desk wasn't for me. Sadily we are always under NDAs, so using a tool like this would require a lot of hoop jumping unless it definitely didn't cache or anything

Really interesting to see this space developing. I'm building a masonry-specific quantity takeoff tool (vision model extraction into a parametric domain model that spits out bid-ready quantities) and the "data prison" framing resonates hard.

One thing I've learned going deep in a single trade: the distance between "structured JSON from a drawing" and "numbers an estimator will bid with" is enormous. I've been really impressed with Bobyard and SketchDeck especially.

h317's point about the liability-driven re-counting circus is spot on. Each party in the chain needs to own their numbers. Revit could have solved this a long time ago had this not been the case. An API that makes each individual count faster is valuable but it doesn't collapse the chain.

Would love to talk to anyone else building in this space.

Looks cool! Where are you getting the data to finetune the cv models for element extraction? I'm worried there isn't a robust enough dataset to be able to build a detection model that will generalize to all of the slightly different standards each discipline (and each firm for that matter) use.

good q — we don't train on customer drawings. Our detection models are trained on a curated dataset of architectural drawings we've sourced and labeled ourselves, focused on the most common fixture and element types across CSI divisions.

The generalization problem you're pointing at is real and it's the hardest part of this. Our approach is to keep the detection scope tight — rather than trying to generalize across every firm's conventions, we train on a small but high-quality set of fixtures and optimize for precision within that scope.

The result is high confidence outputs on the elements we support, rather than mediocre coverage across everything.

We're expanding the detection surface incrementally as we validate accuracy division by division!

How in the world is an answer to a question from the account posting TFA replying directly to said question getting killed?

When will this be available for 30000x8000px electrical diagrams?

I have to make a BOM and oh boy I hate my job

What software made the bitmap? Seems like a step earlier in the pipeline could help generate a BOM more easily.

I'm not really sure and I don't have access to it, I just recive flat PDFs or TIFFs

A lot of them are "archival" so I'm pretty OOL

You might even be SOL

It is telling that so many of the comments here assume the person with a thing that is not the most practical would be easily able to request thing in a different format. The assumption that the person with the inconvenient thing would never have thought to ask if more convenient thing was available and just willfully toiling with the inconvenient thing is kind of insulting.

Also, in the construction industry you get an updated drawing file a day before the bidding closes... good luck getting the GC to send more detailed files (that they themselves got elsewhere) in that time. You're better off sending it to your estimation department in India and letting them work through the night to put together the new estimations.

I’m building a similar platform, with electrical being furthest ahead - SLD, panels, lights, power, comms.

Also do doors, windows, and mechanical equipment.

dm, and I can include you in the next preview.

I work in the automotive field, I don't know if this complicates the things further but I appreciate any help!

I'm not sure how to dm on here, but I'm very interested

You can paste "who is alexeischiopu" to a search engine, and since there isn't an athlete with the same name, a good candidate appears.

Can reach me at alexei@usevawn.com

What do you hate the most?

silly questions

Doesn't Gemini do this directly? I uploaded an image and asked it to identify the doors and it gave me a JSON array of the boxes.

«Why we did it»; would rather have a “How we did it”. The why gave me AI generated marketing material feelings.

Tailscale’s article about NAT traversal is an example of how to write “how we did it”: https://tailscale.com/blog/how-nat-traversal-works

Fair point and thanks for the tip. No AI here though.

I know its been a few days, but have you looked into the new Bluebeam offerings and do you see that as competition for you guys?

https://www.bluebeam.com/bluebeam-max/

Their first example is counting fixtures

Fair callout — Bluebeam Max does touch estimation. That said, we've talked with 20+ subs and almost none know them know exists.

The reality is people that work in construction (Excluding field work), refers to AI as a scam and compares it to the crypto hype back in 18 .. compares it to crypto in 2018 — all hype, no landing.. seriously they do not like it.

I built this selfishly. Eight months ago I wished I had a programmatic interface to interact with drawings. So I built one.

last point: our customers aren't just doing takeoffs and they aren't just subs or gc's either — they're building tools in the construction tech space and estimation is one use case. The broader opportunity is making all that structured data accessible.

great q!!

Also I didn't see Claude there "ai" tool (a bit weird they are letting the model providers own the entire interface) count any fixtures, and if it did it was probably pulling it from the markup's table which is enabled via their api, which you can technically just slap an llm on top of it. But hey kudos to them - this space is not easy. (https://developers.bluebeam.com/s/studio?language=en_US)

I have been working on an extension of this problem lately that involves extracting all doors + any details about those doors to produce quotes. I have found giving the pdf to codex pretty good at it as it can take subcrops of the plans to look at certain areas of high noise in more detail. Only downside is cost is quite high.

Very interesting. Im on vacation but will check this out at work next week.

What is the maximum resolution you support for PDFs? The max gemini will do is 3072x3072. We have plans that are 10x that size.

What divisions information do you need?

Mainly 09, but also 05 and 07.

Won’t get to that soon. Focusing on MEP with some architectural elements).

Interesting, I've dealt with a similar problem before. If anyone needs a self-hosted/custom solution, feel free to reach out.

Do that for Finnish construction documents. My parser is 30000+ lines candidate based but the lack of standards and the Finnish language...

Have you tried using it with Finnish construction documents? It should work for the detections that cover drawings.

Love it! Starbucks Vente Machiato sip

Love to give it to an arc client, not sure who the right person to implement this would be? Hmm…

Hey OP here - Love to help if you're looking for a team to implement a solution.

https://cal.com/anchorgrid/anchorgrid-external-meeting?durat...

How can this tool help easy my workload? I'm already pretty overwhelmed.

This is incredible

What do you foresee being the end use case for this (or most valuable use case)?

Anyone building in or for construction tech — whether that's a startup building estimating or project management software, a construction company with an internal tech team solving this themselves, or a builder looking to automate their workflow. The common thread is drawings. Every one of those groups lives and dies by their ability to extract actionable data from a PDF that was never designed to be machine-readable. We're building the layer that makes that possible so they don't have to start from scratch.

Why does the workflow lie at the level of a real or virtual piece of paper and not in the metadata from the applications used to create that piece of paper? Seems like a CAD tool would allow you to identify each element of the drawing, assigning metadata as required.

Only a small set of construction stakeholders participate in the CAD ecosystem (e.g., architects, large GCs) while a broader set of stakeholders (subcontractors, trades, smaller GCs/CMs) do not receive BIM files and work with PDFs. CAD/BIM is a wonderful aspiration but for many the reality is PDFs.

Re. "CAD/BIM", technically speaking CAD doesn't imply BIM, and the industry's promotion of BIM is akin to AI promotion among software engineering teams - the benefits aren't clear upon detailed review of the advertised capabilities. The CAD part, on the other hand, is generally recognized as the essential tooling for the profession and I'm surprised to hear that it just is a "wonderful aspiration".

"The profession" actually is a wide variety of trades, not just architects and contractors. Electricians, plumbers etc. where CAD is not yet widely spread. Which hopefully will change in the near future, with open source BIM tool chains, boosted by generative/agentic AI.. Finally, a huge source of confusion and execution hiccups will be overcome.

Until then pdf rules!!!

Oh you sweet summer child. These draws are anywhere from 0 to 120 years old and might just be something pulled out of a floppy disk from 1970 to scanned in coffee ridden pieces of paper sitting in a desk folded a hundred times.

The world in which metadata is a common thing attached to any file doesn't exist, and probably never will, no matter how much you try to improve CAD work flow.

> Oh you sweet summer child

I know you're just repeating a phrase from a TV show but do you know how incredibly condescending this comes across to most people?

Great points raised!

Good idea :)

Thanks!!

cool. What's pricing like?

Thanks! https://www.getanchorgrid.com/developer/pricing

Let me know if you find it useful or have any questions, happy to help.

Thanks -- btw the Pricing link on the site pulls up a form, not that page.

Your smart features looks like a game changer? Nice job!

Thanks!!

Of course it is not working. PDF and images are supposed to be tamper resistant. OCR tries to reverse engineer them.

Since when is tamper resistance a part of PDF or any common image format?

PDF files can be signed, that is tamper resistance. Tamper resistance doesn't have to make any difference to the readability of the document.

So can any type of file -- that doesn't have any relevance to the supposed design of every file type in existence. Now, later versions of PDF do have explicit support for signatures, but what does this have to do with preventing OCR? OCR reads a file, it doesn't change the original file.

Some OCR solutions do change the original file, like OCRmyPDF. They take layers that were just images before and replace it with text layers so that you can search the document.

That isn't OCR, but an application of the resulting output of OCR. Again, a signature on a PDF or any type of file doesn't prevent you from reading it. (It also doesn't technically prevent you from changing it, it just enables the detection of changes to a particular file.)

There's nothing about PDFs or image formats that prevent anyone from doing OCR. The reason construction documents are difficult to OCR is because OCR models are not well trained for them, and they're very technical documents where small details are significant. It doesn't have anything to do with the file format

True but you can make modified copies if you reverse engineer it with OCR.

That's not really what I would call reverse engineering. If you read a pdf, and type it into word is that reverse engineering? Either way whatever you get is in no way going to convince anybody that it is the original.

Can't one just remove the signature and re-sign it with anything else after tampering? Who verifies PDFs that hard?

If you're performing OCR, you're almost by definition, disregarding the source file. The whole point of OCR is to be transformative.

You can't change a PDF, it is by design to be not easy to OCRed

PDFs are merely an collection of objects, that can be plainly read by reading the file -- some of those are straight up plain text that doesn't even need to be OCR'd, it can be simply extracted. It is also possible to embed image objects in PDFs, (this is common for scanned files) which might be what you are thinking of. But this is not a design feature of PDF, but rather the output format of a scanner: an image. Editing PDFs is a simple matter of simply editing a file, which you can do plainly as you would any other.

It is not by design! PDFs that are made from scanned documents or collections of images would require OCRing but that is true of any format that the scans/images are put into. These days the vast majority of PDFs do not need to be OCRed as the pages are just made up of text, line drawings and images. And although it can get tricky you can edit those text, line and image commands as much as you want.

For example: add this is in the contents stream for a pdf page and it'll put hello world on the page

  BT
    /myfont 50 Tf
    100 200 Td
    (Hello World) Tj
  ET

(Note: a bit more is required to select the font etc)

Completely false.

Reply on news.ycombinator.com