Image Embedding

Read page and visual-object images from a document visual manifest, generate image vectors, and write them to an image index table.

Use this node for visual retrieval over drawings, layouts, and image objects. Input should come from Build Document Visual Manifest.

Prepare index inputs

Templates bind the manifest and index configuration. When building a flow yourself, provide the following inputs.

Input

Description

Visual manifest

Manifest object or file reference containing readable page and object image references

image_vector_table

Target table for image vectors

image_embedding_model

Model that processes actual images

embedding_dimension

Positive dimension matching both model output and target index

preprocess_version

Required identifier for image preprocessing

distance_metric

Required metric matching the image index configuration

The node is enabled by default. Disabling it returns disabled and skips embedding and index writes. Set the embedding source to real to process the images themselves.

Check index coverage

written counts all rows written. page_rows and visual_object_rows count page and object records separately.

all_source_file_ids lists the sources considered; indexed_source_file_ids lists only sources actually written to the image index. Inspect file_statuses for each file’s result.

In flows that allow empty image results, no_indexable_images means the file had no images to index and no image vectors were written for it.

Example: index drawing images

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

Pass the full manifest and index settings. After running, check dimensions, row counts, and per-file status. Pass the indexed source list to Lineage Registration to record source relationships.

Last updated on