CAD Drawing Segmentation

Analyze drawing pages to identify regions, text, and part relationships, producing drawing reports or retrieval units.

Input drawing PDFs or images for visual retrieval, information extraction, and indexing.

Configure input and mode

Setting

Description

Files

Catalog file_ids of drawings to process

LLM model

Workspace model for analysis

Vision model

Model for OCR and visual assessment

Parsing mode

Prefilled: deep; fast uses coarse retrieval parsing

Report mode

Prefilled: drawing_report; segments returns segmentation results, and retrieval_index supports fast retrieval

CAD segmentation service URL

Optional self-hosted endpoint; leave empty for the built-in shared service

Fast mode uses retrieval_index. Deep mode uses drawing_report or segments.

Control the processing budget

Configuration defaults are 16 agent rounds, 4 concurrent tools, 64 candidate regions per page, 24 crops, 18 refinements, and 12 visual assessments.

The default CV profile is cad-default-v1, the SAM profile is sam2-default-v1, and the artifact level is review.

In drawing_report mode, set maximum agent rounds to 0 to skip additional agent enhancement and retain the base results.

Use the output

Segmentation or report paths can return segments, parts, relationships, and ocr_texts. Fast retrieval returns units such as items. Read the fields corresponding to the chosen report mode.

manifest_file_id identifies the result manifest; artifact_files contains related artifacts. Check status and review for result status. Retain the full result, especially page_sources and image references, when building a visual manifest.

Example: build a visual drawing index

Flow: Read MOI Volume → CAD Drawing Segmentation → Build Document Visual Manifest → Image Embedding.

Choose deep parsing and drawing report mode. Pass the full CAD result to the manifest node’s cad_result, then use the manifest to build the image index.

Last updated on