Digitising a paper drawing archive with AI: title blocks, revisions and similar drawings

What digitising a drawing archive with AI involves

Digitising a paper drawing archive with AI means scanning every sheet and letting a vision-language model read each title block and parts list into fields such as drawing number, revision, title and material; software then builds the revision chains, links each sheet to its ERP part number and makes the archive searchable by text and by similarity of the drawing. By our estimate one server with four NVIDIA L40S cards reads the title blocks of a 60,000-sheet archive in about a working day, and the drawings, which carry your design knowledge, stay on that server.

This guide is for the owner, the managing director or the head of engineering or document control at a manufacturer with decades of drawings in plan chests, folders or on microfilm. It explains each step in plain words and keeps the hardware in one section. Our article on AI quoting from customer drawings covers new incoming drawings; this one covers the archive you already have. Our guide to an AI server for archives and libraries covers historical handwriting in public archives, which is a different job.

The reasons to do it are familiar. A service technician needs the drawing of a spare part for a machine built in 1994. A designer wants to know whether a similar bracket already exists before drawing a new one. An auditor asks for the revision that was valid when a batch was made. In a paper archive each of these is a search by hand, and it depends on the people who know where things are.

Scanning the drawing archive: what it holds, sheet sizes, resolution

A typical archive mixes sheet sizes from A4 to A0, original drawings and blueprints, stamps, handwritten changes and notes added in the workshop. The part that matters most for indexing is the title block, the box usually placed in the lower right corner of the sheet. ISO 7200, last confirmed in 2025, "specifies the data fields used in the title blocks and headers of technical product documents", with their names, contents and lengths. Older drawings in your archive follow whatever your company's design office used at the time, so the first task is to list the fields your title blocks carry across the decades.

Sheet sizes per EN ISO 5457 as listed by Roymech, read 25 October 2026; pixels and file sizes are our arithmetic for 8-bit greyscale without compression. Black-and-white (1-bit) scans take an eighth of these sizes before compression.

Scan resolution is the first decision. This article plans with 300 dpi (dots per inch); faint pencil originals and very small figures may need more, which a few test scans show. Whether you scan yourself on a large-format scanner or use a scanning service, agree the resolution, the colour mode and the file format before the first sheet goes in, and keep the original scans unaltered: every AI step works on copies.

Title block extraction: reading title blocks and parts lists with AI

A vision-language model is an AI model that looks at an image and writes what it sees as text. Asked for the fields of a title block, it returns them in a fixed structure that a database can take directly: drawing number, revision index, title, material, scale, date, drawn by, approved by.

Qwen3.8-27B is an open model that its model card describes as "a native vision-language model that understands images and videos", under Apache 2.0. The model card reports document results such as 91.1 on OmniDocBench 1.5 and does not mention technical drawings, so its accuracy on your sheets has to be tested. Its image processor makes one image token, a unit of input, per block of 32 × 32 pixels and accepts at most 16,777,216 pixels per image.

That limit is why the model never sees a whole large sheet. An A0 scan at 300 dpi would have to be shrunk to about 104 dpi to fit, by our arithmetic, and small text would become unreadable. The pipeline therefore cuts out the regions it needs at full resolution: the title block, the parts list and the revision table, and later each view for similarity search. A title block of 190 × 70 mm, our example size, becomes 2,244 × 827 pixels at 300 dpi and about 1,800 image tokens. Finding the title block on a sheet is a simple step, because it sits in the same place on most sheets of one design office.

Handwriting and poor copies. ki-syndikat.de, a German catalogue of AI use cases for small and medium-sized companies, reports in its plastics section that printed and cleanly scanned documents reach OCR accuracies "von 95 Prozent und mehr" and that handwritten production cards from the 1990s and 2000s fall back to 60 to 75 per cent when poor scan quality and handwriting meet. It adds that everything extracted must be checked by a specialist before it counts as valid data. Drawings have the same problem in their revision notes and workshop annotations. Plan a review step: the AI fills the fields, a person confirms them, and every value keeps a link to the cut-out it came from.

A test before the project. Take 200 sheets from different decades and design offices, including the worst copies. Have someone from document control enter the correct title-block fields once, run the model, and count the errors per field. The result tells you which fields the AI can fill alone and which need a person every time.

Revisions, duplicates and the link to the ERP

Once each sheet has its fields, ordinary software does the rest, with no GPU. Sheets with the same drawing number form a revision chain, ordered by revision index and date; gaps and two sheets claiming the same revision become a list for document control to resolve. Exact duplicates, such as an original and its blueprint, are recognised by their identical fields and filed together.

The link to the ERP turns the archive into a working tool. Where the drawing number is the part number, or a cross-reference table exists, each sheet is attached to its article in the ERP master data, so a buyer or service technician opens the drawing from the part. Where old numbering schemes differ from today's, the matching rules come from the people who know the history, and the AI suggests candidates for them to confirm. Our guide to order entry from emails and PDFs into the ERP describes the same pattern of AI proposal and human confirmation for orders.

We supply servers with L40S or RTX PRO 6000 Server Edition cards, and our Private AI/ML service sets up the reading pipeline and search; send us twenty typical sheets and the size of your archive, and we reply with a configuration and a test plan.

Find similar drawings with AI: search by image

Designers often search for "something like this part" without knowing a number. An image model can do this search. It turns each view of a drawing into an embedding, a list of numbers that describes the picture so that similar pictures get similar numbers, and the search compares the numbers of your query with those of the archive.

Google's SigLIP 2 is one open option. Its model card, under Apache 2.0, names "zero-shot image classification and image-text retrieval" as uses; the so400m version takes images of 384 × 384 pixels and returns vectors of 1,152 numbers, and its file is 4.54 GB. Because it is made for image-text retrieval, it can also find drawings from a typed description such as "flange with six bolt holes". Each view is scaled down to 384 × 384 for this, which keeps the shape and loses the dimensions, so the result is a shortlist that a person opens and checks. How well it separates technical drawings is not published; test it on views from your own archive.

The index is small. At 1,152 numbers of 4 bytes, one view takes 4.6 KB, so 180,000 views, three per sheet for 60,000 sheets, take about 0.8 GB by our arithmetic. Our guide to vector databases for RAG compares where to store them, and our article on ColPali for visual document retrieval covers searching text-heavy pages by image.

The hardware: which server, how fast, what the building needs

Reading title blocks is a batch job: the server works through the queue of scans at night or over a weekend, and nobody waits for an answer. By our users-per-card rule, 90 per cent of the memory the driver reports less 3 GiB per card, one 48 GB L40S with ECC on leaves 37.3 GiB usable, as our guide to buying an L40S server derives. Qwen3.8-27B in FP8 takes 28.7 GiB of it, and a title block with its prompt and answer takes about 0.21 GiB with an FP8 cache, including the model's fixed state per request, so one card holds about 39 sheets in flight by our estimate.

For speed we use the closest published measurement. StorageReview tested Qwen3.8-27B in an NVFP4 build on an L40S with vLLM 0.27.1 (14 October 2026): with 32 requests at once, 8,192 tokens in and 1,024 out, the card produced 190 output tokens per second in total. A title block needs about 2,100 tokens in and, by our assumption, 300 out, so by our estimate one card reads in the order of 2,300 title blocks an hour. The FP8 model may run at a different speed on the L40S, so a test on your sheets replaces this figure.

Our estimates: memory by our users-per-card rule with one copy of Qwen3.8-27B FP8 per card and an FP8 cache; speed scaled from StorageReview's L40S result (190 output tokens per second at 32 requests, 14 October 2026) at 300 output tokens per title block; power is board power (350 W per L40S, 600 W per RTX PRO 6000 Server Edition) plus two processors plus 15 per cent, as on our L40S and AI server pages.

What this means for a project. An archive of 60,000 sheets, our example, needs about 26 card-hours for the title blocks by this estimate: about 6.5 hours on a four-card server, or about 13 on a two-card one. Parts lists and revision tables add work in proportion to their length, and the similarity index adds a small pass over all views. Compare these figures with the number of sheets your scanner or scanning service delivers per day. The server only has to keep up with the scanning, and by this estimate a two-card server keeps up with several thousand sheets an hour.

Which card. The NVIDIA L40S is a passive 48 GB data-centre card at 350 W for rack servers; a two-card server runs on one 230 V 16 A circuit and belongs in a server room or a ventilated rack. The RTX PRO 6000 Blackwell Server Edition has twice the memory and holds about 252 sheets per card by our estimate; it pays off when the same server will also run the company assistant, document search or AI quoting from drawings after the archive is done. Our comparison of the L40S and the RTX PRO 6000 Server Edition sets out the differences.

Notes for the IT team

1. Storage. Keep the original scans on the file server or the document management system with a backup; the pipeline reads copies and writes cut-outs, fields and vectors next to them. At 300 dpi greyscale, uncompressed A0 sheets take about 140 MB each, so plan capacity from the format mix and the compression you choose.
2. Serving. Run the model with vLLM, switch thinking off per request with enable_thinking set to False in chat_template_kwargs, as the Qwen3.8-27B model card describes, and force the answer into a JSON schema for the title-block fields.
3. Ada and FP8. The L40S is an Ada-generation card (compute capability 8.9). How Qwen's block-wise FP8 checkpoint runs on it depends on the engine; our article on Qwen3.8-27B on the L40S covers the options.
4. Search. Store the SigLIP 2 vectors in a vector database next to the title-block fields, so that a search can combine "similar shape" with "material steel" or "drawn before 1990".
5. Access. Drawings carry the design knowledge of the company; give the archive the same access rights as the CAD vault and log who opens what.
6. Quality. Keep the 200-sheet test set and rerun it after every model or prompt change.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the L40S and the RTX PRO 6000 Server Edition and build servers with two or four cards to order, assembled and tested before delivery, on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the reading pipeline, the review screen and the search; describe your archive, sheet sizes and ERP, and we propose a test with your own drawings.

FAQ