Enterprise document intelligence teams now have a lightweight tool that spots every image in a PDF without touching the pixels. The new companion to the parsing series feeds a table called image_df that lists page, bounding box, size and a hash for each picture. Knowing where every photo sits is the first step toward searchable content.

The raw list alone does nothing for retrieval; a bounding box offers no searchable text. Instead, the system runs a cost‑ordered cascade: a cheap filter removes tiny or repetitive graphics, a type check flags plain panels, OCR reads clean tables, and a vision model interprets charts and photos.

Key to efficiency is a lightweight filter that discards icons, rules, and logos repeating across pages. It keeps only images that may hold meaning, then classifies each into text, chart, or photo. The decision tree ensures OCR is called only on pure text, while costly vision calls surface the remaining visual data.

By prioritizing which images to process, the framework saves on expensive model calls and reduces latency in retrieval‑augmented generation pipelines. The result is a more responsive system that still delivers rich, searchable depictions of figures and diagrams without paying to read every pixel.