Frequently asked questions

Windows 10 and Windows 11. The installer runs directly — no Python, CUDA or other runtime setup. CPU detection works; with a discrete GPU you can switch to GPU inference.

See the system requirements listed on the download page. Detection runs locally and works in CPU mode; for high volume, a machine with a discrete GPU is recommended so you can enable GPU inference.

First make sure the installer came from the official download page. If it still will not start, send us the version number, your Windows build and a screenshot of the message by email or WeChat, and we will work from the actual error.

Local detection, review, numbering and export work offline. Sign-in, quota lookup, updates and cloud OCR require a network connection.

Download the latest installer and install over the existing version. Release notes are on the release notes page. Updating does not reset drawings or export files already on your machine.

We do not put a blanket percentage on the website. Results depend on drawing clarity, annotation density and annotation style: common mechanical dimension annotations are stable, while scans, handwritten edits and dense annotations will need review. Detection uses a YOLO vision model, and both model and confidence can be adjusted in the client; review and confidence filtering exist for exactly those cases.

A typical mechanical drawing takes about one second to detect (local CPU / GPU inference in the client). Very large drawings, very dense annotations or slower machines take a little longer; the client status bar shows the actual time.

Yes. OCR results are reviewed item by item, values can be corrected and low-confidence items filtered by threshold. Corrected results can be saved and used as training samples to improve the model.

Dimension detection runs locally on a YOLO vision model with CPU or GPU inference; drawings do not have to be uploaded first. Only when you explicitly choose cloud OCR for a job does the related image leave your machine; cloud recognition currently uses the DeepSeek vision model.

Detection itself is measured in seconds; total time depends on page count, annotation density and whether GPU inference is used. The honest answer is to run your most common drawing through the free quota and judge from that.

Each OCR result carries a confidence value. You can set a threshold so low-confidence items surface first for review instead of checking everything from scratch.

The product targets dimension and tolerance annotations in mechanical drawings, and generates balloon numbers and inspection details from them. Annotation styles differ a lot between industries, so validate coverage with your own drawings; when needed we can assess a custom model.

PDF, PNG and JPG. Export CAD-native files to PDF or images before importing them.

Pages are managed individually: process one page or run batch detection, numbering and export across a whole set, with results saved per page.

They can be imported, but clarity, skew and glare affect detection and OCR stability. For important deliverables use an exported PDF or a high-resolution image; if a scan is all you have, cloud OCR can improve extraction stability.

Use a clean PDF export or an image with enough resolution. Heavily compressed or blurry images reduce OCR quality and increase the review workload.

A ballooned PDF (numbers and leader lines), an Excel detail sheet (index, inspection item, feature description, nominal, tolerance, measured value — GB/T 1804 as the default tolerance) and an inspection report (templates and field mapping). All three can be produced in one export.

CAD-native formats are not imported directly today; export to PDF or an image first. If your drawings only circulate natively, tell us and we will assess support.

The free trial includes 200 detections. Monthly membership is ¥49 per month and annual is ¥490 per year. Team licensing and private deployment are quoted on actual scale.

Not yet. Create an account and use the free quota; when you are ready to activate, contact us by phone, WeChat or email. A person confirms the account and usage scale before activation.

Per drawing page: one complete detection run consumes one detection. OCR review and export are not charged again. You can check the quota on your account page.

Yes, standard VAT invoices. Provide the invoice title and tax number when activating; the exact process is confirmed by email.

Software is a virtual good, so activated memberships are not refundable in principle. Validate the full workflow with the 200 free detections before you decide.

Quoted per device count, concurrency and data requirements. Intranet deployment and offline models can be assessed. Tell us roughly how many devices and which drawing types, and we will send a concrete plan.

Not by default. Preview, dimension detection, ballooning and export all run locally. Data leaves your machine only when you explicitly choose cloud OCR. The full description is on the data & security page.

Yes. Cloud OCR runs on the DeepSeek vision model and is metered per run; the scope is visible before a job starts and usage can be checked in your account. If you do not choose it, no such charge is generated.