Back to Blog
AI

Designing the Same Part From Scratch Again: Raising Design Reuse With AI-Powered Similar Drawing Search

Searching drawings by file name and part number leads to duplicate designs and unnecessary tooling and inventory costs. Here's how AI-powered similar drawing search, combining shape embeddings with attribute filters and tied into PLM and ERP workflows, raises design reuse.

POLYGLOTSOFT Tech Team2026-10-267 min read1
Similar Drawing SearchCADDesign ReuseImage EmbeddingEngineering AI

Tens of Thousands of Drawings, Found Only by Part Number and Memory

Engineering teams at manufacturers often sit on tens of thousands of drawings. Yet when a new part is needed, the way people look for an existing one usually comes down to file names, part numbers, and what a senior designer happens to remember. If the numbering scheme has changed or that designer has left, someone ends up redrawing a bracket that is nearly identical to one already on file.

Duplicate design costs far more than one extra drawing:

  • Tooling and fixtures: Parts that differ by a few millimeters each get their own mold or jig
  • Process: Machining conditions, inspection sheets, and NC programs are written and validated all over again
  • Inventory and purchasing: More part numbers mean fragmented safety stock and order quantities, and fewer volume discounts
  • Suppose registering one new part number takes a few dozen hours across design, validation, purchasing, and quality. A hundred part numbers a year that could have been existing parts then add up to thousands of hours. Reusing proven parts, on the other hand, lets you skip prototype and approval steps and shortens development lead time.

    How Similar Drawing Search Works

    Combining Three Signals

  • 2D drawing image embeddings: Drawings are rendered as images and turned into feature vectors by a vision model, so similar shapes can be found by vector distance
  • 3D shape features: When native 3D files such as STEP exist, volume, surface area, feature composition (holes, pockets), and shape descriptors can be extracted and compared. This is more accurate than 2D alone
  • Title block and note text: OCR plus a language model reads part name, material, heat treatment, surface finish, and tolerance notes into structured fields
  • Hybrid Search

    A part that looks similar isn't necessarily one you can use. What works in practice is to retrieve candidates by shape similarity, then narrow them with attribute filters such as material (SUS304, AL6061, and so on), overall dimension ranges, and finish. A designer uploads a sketch or an existing drawing and adds conditions like "same material, length within ±10%."

    Data Preparation Decides the Outcome

    Real drawing repositories mix scanned paper, print-ready PDFs, and native CAD files like DWG and STEP. The first job is normalization: fixing resolution and orientation, separating the title block and border, and linking the different formats of the same part.

  • Revision history: If Rev. A through D of the same part all show up, results get cluttered. Keep the latest approved revision as the representative and group the history beneath it
  • Obsolete drawings: Don't hide them; flag their status so designers can see why they shouldn't be reused
  • Title block cleanup: Values like "SUS304," "STS304," and other informal spellings must map to a standard code, or attribute filters won't work
  • It Only Gets Used When Connected to PLM and ERP

    If search results show only a picture, designers still have to dig through ERP afterward. Each result card should show part number, latest unit price, supplier, usage over the past year, and where-used products, so the "let's reuse this" decision can happen on the spot.

    Even better is to build it into the workflow. Attach similar-part results automatically to every new part number request, and when a close match exists, require a reason for not using it. That turns search from a matter of individual goodwill into a checkpoint in the approval process.

    Security and Access Control

    Drawings are core technical assets, and customer drawings are often covered by NDAs. Sending originals to an external cloud API may be ruled out by contract, so on-premises deployment of the embedding model and vector database should be the first option evaluated.

    To keep the search system from becoming a leak path, access rights by customer and project must mirror the PLM permission model. Parts exclusive to Customer A shouldn't appear in search results for a designer working on Customer B's project, and views and downloads should be logged.

    Pilot Scope and POLYGLOTSOFT's Approach

    Rather than tackling every drawing at once, start with one part family in one product line, such as sheet-metal brackets or shafts. Useful metrics include:

  • Search hit rate: How often a part the designers already know exists shows up in the top 10 results
  • Reduction in new part numbers: New registrations and reuse conversions over comparable periods before and during the pilot
  • Search time: How long it takes to confirm a similar part compared with the old way
  • POLYGLOTSOFT helps build similar drawing search on top of your existing drawing repository, PLM, and ERP without replacing them. We work with you step by step, from drawing normalization and title block cleanup to hybrid search, integration with new part registration, on-premises deployment, and permission design. If repeated duplicate design is costing you time and money, let's start by scoping a pilot together.

    Need Technical Consultation?

    Our expert consultants in smart factory, AI, and logistics automation will analyze your requirements.

    Request Free Consultation