Strategies

Operator sets strategy=auto for a mixed corpus of clean digital PDFs and scanned images, expecting Unstructured to pick per document.

auto routes per document/page between text extraction and OCR/layout. Use auto when the corpus is heterogeneous; do not assume auto always equals hi_res or always equals fast — verify behavior by inspecting which elements carry coordinates (layout) vs plain text.

To cut cost the agent forces strategy=fast for every document, including scanned invoices and image-only PDFs.

fast is text-only (no OCR, no layout model) — it works for born-digital docs but yields empty/garbled output on scans. Reserve fast for documents with an embedded text layer; route image/scanned docs to hi_res or ocr_only.

A document with multi-column layout, figures, and tables must be extracted with accurate element typing and table structure.

Use strategy=hi_res so the layout-detection model assigns element types, table structure (with infer_table_structure), and coordinates. Accept that hi_res is slower/costlier than fast — the tradeoff buys layout fidelity the workload requires.