partition
Subtype: vlm
For more information about how the various partitioning strategies balance output quality, speed, and cost, see Partitioning.
Usage guidance
Use this strategy when documents are consistently image-heavy, scanned, or visually complex and every page should be processed by a vision language model for maximum fidelity. Unlike the Auto strategy, VLM routes every page directly to the vision language model without per-page evaluation.Settings
VLM provider to use. Allowed values:
anthropic, auto, bedrock, openai, vertexai. Default: anthropic.If specified, overrides the default API key used for VLM provider calls. Default: none (uses Unstructured’s internal default).
VLM model to use. Default:
claude-sonnet-4-20250514.For a full list of the models available in Unstructured, see Available models.Format of the partitioner response. Allowed values:
text/html, application/json. Default: text/html.If
true, applies Beautiful Soup’s prettify method to the HTML generated by the VLM partitioner, adding indentation for readability. Default: true.If
true, assigns UUIDs to element IDs, guaranteeing uniqueness (useful when used as database primary keys). If false, assigns a SHA-256 hash of the element’s text as its ID. Default: true.Must be
false to use the VLM strategy.If
true, allows pages to be routed to Fast partitioning as needed for better performance and cost savings. Default: true.
