A department-level knowledge base built from 200+ PDF drawings combines visual, semantic, and full-text retrieval to return the Top 3 matches with evidence.
Zhongding traces back to the Ningguo Sealing Parts Factory founded in 1980 and became China's first listed sealing-components company in 2006. It makes automotive rubber and sealing parts, hydraulic and pneumatic seals and rubber-plastic components, supplying global OEMs across more than 20 countries. FY2024 revenue was RMB 18.85bn.
When a customer sends a picture of the part they want quoted, engineers leaf through hundreds of drawings by hand to find the match. Quotation accuracy then rests on individual experience, because there is no systematic way to compare drawing similarity. Writing the process inspection work instruction for a part is equally manual and slow. And because each business unit manages its own data, what one department has already learned cannot be reused by another.
Zhongding ran a POC on a drawing intelligence workbench built on OmniFabric. Over 200 PDF drawings were sliced, read by VLM and OCR, interpreted by an LLM into structured documents and vectorized into the multi-modal database, with a separate knowledge base per department for isolation. An engineer selects a department and uploads a partial-view image; feature extraction runs VLM, OCR and LLM entity recognition, and hybrid retrieval-image vector, text vector and full-text with weighted scoring-returns the top three most similar drawings with match detail. The same knowledge base drives composition of the process inspection work instruction.
A requirement image returns the closest matches from the drawing library in seconds, replacing manual comparison across hundreds of PDFs.
Similarity scores and match detail give quoting a systematic basis instead of relying on whichever engineer remembers a comparable part.
Process inspection work instructions are generated from the approved drawing in the required format, standardizing documents that were written by hand.