A document parser preserves visual cues agents need to find requirements
In plain words
When a Word or PDF document is turned into text, color or highlighting that marks an important requirement can disappear. This paper introduces a code-based parser that keeps those visual cues and adds plain-language hints so an answering model can use them.
On a task using real policy documents, Claude Opus 4.6 found author-marked requirements more successfully with the formatting-aware text than with Docling or plain text. The result is specific to that task and model.
The research
When a document converter drops color or highlighting but keeps the words, can an answering model still tell which requirements the author marked for attention? That matters to the answer this study measures: whether the model retrieves those marked requirements from the converted document.
The parser carries color and highlighting into text for an answering model, and tests whether that helps it retrieve requirements marked by policy authors.
In real policy documents, can an answering model identify requirements that authors marked when conversion keeps the words but drops the visual marks? The paper tests whether preserving those cues changes retrieval.
Claude Opus 4.6 answered the same questions using different document representations, and a deterministic matcher checked the answers against the original marks. The formatting-aware version scored 0.690 micro F1, above Docling and plain text. This supports a benefit from preserving formatting on this task.
The main result comes from a specific requirement-retrieval task on policy documents, and the labels were not yet checked by human annotators. PDF cue recovery also requires a usable text layer, and typed cues did not help every evaluated model.
Analysis
Deep
- Original paper authors
- Yikai Zhao, Saurabh Pandey, Pradeep Kumar Misra
- Submitted to arXiv
- License
- arXiv non-exclusive