Agentic Document AI deployments carry a two-part requirement. The first is extraction precision: pulling the right values accurately from complex, dense documents—payroll reports, financial filings, audit records, the structured documents that define these workflows. The second is attribution precision: for every value your pipeline extracts, proving exactly where on the page it came from.
Most standard document parsers expose coarse, layout-level bounding boxes that can tell you where a paragraph sat or circle an entire table block. But if you were parsing a high-density financial report with 20 different payroll figures or inventory values tightly packed onto a single text line, hovering over an AI’s citation would awkwardly highlight half the page. That level of precision is "good enough" for basic LLM chunking, but it completely falls short for enterprise fintech applications, compliance reviews, and workflows that require verification.
In strict audit workflows, a broad neighborhood is not verification. When twenty values sit on the same line, a blocky citation cannot tell you which one the agent used. That is not verification.
Today we're announcing Granular Bounding Boxes for LlamaParse: line, word, and cell-level tracking across your entire document.
Precision at Every Layer
When you trigger a parse, you can now request coordinates at three levels of granularity:
- Line Level: Isolate specific sentences or individual text lines within a larger layout paragraph block.
- Word Level: Track the exact coordinate regions down to the individual word token layer.
- Cell Level: Pinpoint the exact visual boundaries for individual table cells, cleanly isolating them even if they contain multiple wrapped text lines.
And one strict rule governs what gets a coordinate: only text that is explicitly on the page. Inferred values, AI summaries, and reconstructed content get no bounding boxes.
How It Works
Activating granular coordinates is entirely opt-in and requires passing a single parameter at job creation:
json
{
"output_options": {
"granular_bboxes": ["word"]
}
}
What this unlocks
Audit-grade citations: When a user clicks an extracted financial figure, your product can jump to the exact source page and highlight it at the granularity level you selected, down to the line, word, or cell.
High-precision redaction: Word-level coordinates let you target specific PII or sensitive values precisely, giving redaction tools the exact boundaries they need without manual page-blocking.
Granular bounding boxes are available now in beta across all paid tiers on LlamaParse. For workflows where attribution accuracy is critical, Agentic Plus runs additional verification rounds to improve precision.
Documentation at developers.llamaindex.ai
