It says forms require preserving hierarchy, associating labels and values with their sections, reading checkbox state, and locating fields with bounding boxes. In one example, its checkbox classifier corrected a VLM’s false reading of a marked W-9 checkbox. The article also reports a digital-form example in which a VLM returned two misplaced boxes while LlamaParse detected the field boxes, and describes box drift on a scanned W-4. LlamaIndex presents its own parser and models as the remedy. These are vendor-reported examples, not independent benchmark results; the article’s central point is that form extraction depends on spatial and semantic structure that a text-only representation can discard.
Raw SignalSep 28, 20261 min read1 source
Why Forms Are Hard to Parse: AI Form Extraction vs. VLMs
In one minute
LlamaIndex published a technical article arguing that general-purpose vision-language models are unreliable for extracting structured data from forms.