A PDF can look perfectly readable while being difficult for software to interpret. A table header may apply to three columns. A price may be crossed out. A footnote may change the meaning of a number. Turning the page into plain text can preserve the words while losing the business meaning.
A useful AI trend in 2026 is the shift from basic text extraction toward document structure and grounding. The goal is not merely to produce a fluent summary. It is to preserve enough evidence that an answer or extracted field can be checked against the original page.
Two developments worth separating
On April 8, Nomic announced a research preview with Muna for on-device layout models. The announcement describes extracting document structure such as tables, figures, headings, and reading order, with examples involving browser-based PDF work and coding-agent skills.
On April 22, LlamaIndex announced its ParseBench collaboration with Kaggle. The evaluation focuses on enterprise-document problems including tables, charts, content faithfulness, semantic formatting, and visual grounding.
One development concerns where a parsing component runs; the other concerns how document-reading quality is measured. Together, they make an important point: buyers should inspect both the data path and the structure of the output, not just ask which language model powers the product.
Layout detection is not full understanding
A layout model may identify the location of a table or the order of text blocks. That does not necessarily mean it has read every cell correctly, reconciled totals, or understood an exception buried in a note.
Similarly, a parser that produces attractive Markdown may still attach a value to the wrong heading. A downstream language model can turn that error into a confident explanation. Better prose is not evidence that the original extraction was correct.
This is why source locations matter. A page number or bounding box lets a reviewer inspect the supporting region. Grounding does not make the field true, but it makes verification more practical.
Example: supplier quotations without blind copying
Consider a small retailer comparing PDF quotations from three suppliers. This is an illustrative use case, not a hands-on evaluation.
The retailer defines a schema: item identifier, quantity, unit, unit price, currency, tax treatment, and delivery terms. Each extracted value must retain a source page and, where available, a highlighted region. A reviewer checks mismatched units and verifies whether discounts apply per item or to the whole order.
The workflow should preserve missing values as missing. If a delivery date is absent, the system should not infer one from a different supplier’s quotation. Any subsequent recommendation should distinguish extracted terms from the retailer’s own assumptions.
Build a parser comparison that reflects your files
Use examples from the formats you actually receive, with permission and appropriate redaction. Include scans, rotated pages, merged table cells, long documents, and revised prices. Keep a few difficult files aside rather than tuning every prompt against the same sample.
Compare field-level accuracy, source traceability, correction effort, processing time, and the cost of retries. An overall score can hide a parser that gets names right but repeatedly misreads quantities. For payment-related work, that failure pattern matters more than elegant formatting.
ParseBench is useful evidence about evaluation design, but LlamaIndex also sells parsing products. Treat its announcement as a vendor-authored description, examine methodology, and avoid converting a leaderboard position into a guarantee for your document mix.
Local parsing does not make an entire app local
Nomic labels its release a research preview. Its on-device layout approach does not establish that every downstream reasoning step in a third-party application stays offline. Ask whether page images, extracted text, embeddings, and logs leave the device separately.
Watch for end-to-end evaluations that follow errors from page parsing into final business decisions. The immediate buying opportunity is straightforward: favor tools that preserve structure, show evidence, and admit uncertainty. A parser that flags a questionable cell for review may be more valuable than one that silently fills every field.
Sources
- Nomic: on-device layout research preview — 2026-04-08; accessed November 7, 2026.
- LlamaIndex: ParseBench and Kaggle announcement — 2026-04-22; accessed November 7, 2026.
