renue Adds Pre-reference Function for Part Information to "Drawing Agent", Improving 3D Model Generation Accuracy with Knowledge Base

renue Co., Ltd. (Headquarters: Minato-ku, Tokyo, Representative: Yusuke Yamamoto) has added three new features to its web application "Drawing Agent," which autonomously generates 3D models from 2D drawings: STEP format output editable in CAD software, OCR automatic dimension cross-checking, and an industrial parts knowledge base. This enhances both CAD data compatibility and generation accuracy required in manufacturing sites.

## Update Summary
Drawing Agent is a web application that automatically generates 3D models simply by uploading 2D drawing images.

Since its initial release in March 2026, multiple manufacturing companies have requested the following:

* Output in STEP format that can be edited with CAD software
* Support for PDF drawings
* Improved dimensional accuracy

This update adds the following three new features:

* Complete migration of the 3D generation engine to CadQuery, enabling output in STEP format that can be edited in CAD software.
* Implementation of an automatic cross-checking pipeline for dimension text extraction via OCR and actual measured values of the 3D model.
* Pre-loaded with an industrial parts knowledge base of 10 categories and 100 parts, allowing the AI agent to dynamically reference optimal generation patterns.

As a result, generated 3D data can now be directly imported into CAD software for editing.
Support for PDF format drawing input has also been added.

## Related Release
In the initial release of Drawing Agent (March 2026), the basic function of autonomously generating 3D models from 2D drawing images was launched.
https://prtimes.jp/main/html/rd/p/000000025.000091210.html

## New Feature Details
### 3D Model Output in STEP Format
The previous release supported output in GLB and STL formats. While suitable for web viewers and 3D printing, these are mesh formats, making it impossible to add faces or change dimensions in CAD software.

This time, the 3D generation engine has been completely migrated to a CadQuery base.

* CadQuery builds solid models using B-Rep (Boundary Representation), enabling output in STEP format compatible with major CAD software for editing.
* CadQuery builds solid models using B-Rep (Boundary Representation), enabling output in STEP format compatible with major CAD software for editing.
* Generated STEP files automatically pass re-reading tests, face count verification, and watertightness checks using Open CASCADE Technology (OCCT).
* Simultaneous output with GLB and STL continues to be supported.

### OCR Dimension Extraction and Automatic Cross-Checking
Quality verification in the previous release primarily focused on visual comparison using Gemini Vision. There was a challenge in detecting cases where "the appearance is similar, but the dimensions differ."

1. An preprocessing pipeline has been added to extract dimension text from drawing images using OCR and pass it as numerical data to the AI agent. Unlike the method where the AI "visually" reads dimensions from the image, this provides accurate numerical values.
2. The actual measured values of the generated 3D model are automatically cross-checked against the dimensions from the original drawing. If the tolerance is exceeded, it enters a regeneration loop, improving quality over a maximum of 10 iterations.

By combining visual comparison and numerical dimension comparison, a multi-layered verification process has achieved an average score of 81/100 for 5 part types.

### Industrial Parts Knowledge Base
An industrial parts knowledge base covering 10 categories and 100 parts has been pre-loaded.

* Cooling equipment, power supplies, servo drives, connectors, enclosures
* Mechanical parts, sensors, pneumatic equipment, structural components, fasteners

Each part information includes CadQuery code generation patterns. The AI agent determines the part type from the drawing and dynamically references the optimal pattern using the `list_known_parts` and `get_part_knowledge` tools.

Part information is managed in JSON format, eliminating the need for code changes when adding new part types.

## Changes Since Last Release
### Changes from Previous Release

* Output Format: GLB/STL only → **Added STEP format** (directly editable in CAD software)
* Quality Verification: Visual comparison only → **OCR dimension extraction + numerical cross-checking** (multi-layered verification)
* Generation Accuracy: Improved accuracy confirmed through iterative refinement
* Input Format: Images only → **Supports PDF input**
* Knowledge Base: None → **Pre-loaded with 10 categories and 100 industrial parts**

### Use Cases

* Importing 2D drawings received from overseas branches into the head office's CAD environment as 3D data.
* Batch converting legacy paper drawings (in PDF format) into 3D CAD data.
* Preparing 3D data for assembly procedure manuals or structural evaluation without the need for CAD operators.

## Future Outlook
The company plans to expand the parts information database to approximately 500 items by the end of fiscal year 2026, strengthening support for high-difficulty categories such as automotive interior parts and rubber molded products.

Automated generation of multi-part assemblies, authentication and multi-user functions, and API integration with PLM/ERP systems are also planned for phased release.
https://youtu.be/7PII_P8l7gU

## Company Profile
Company Name: renue Co., Ltd.
Location: 5F, Shiodome City Center, 1-5-2 Higashi-Shimbashi, Minato-ku, Tokyo, 105-7105 Japan
Representative: Yusuke Yamamoto
Business Description: AI consulting business
URL: https://renue.co.jp/

## Contact Regarding This Matter
Email: info@renue.co.jp