We've built the most accurate document extraction system on the planet. Extract Max Effort will spend dynamic tokens reasoning about every visual detail across hundreds of pages to ensure that you get the maximum accuracy possible. Check out this massive well log report below. It's more accurate and still... cheaper than the frontier agent harnesses on max thinking! (astra xhigh, opus 5.5 xhigh). Check it out: LlamaParse:show more

@JSON_JEFF

can i cap the spend per document?

What decides whether this ships isn't accuracy, it's whether it flags which fields it's unsure of. At 98% on 500 pages you still get ~10 wrong values; unmarked, someone re-reads all 500. Does Max Effort emit per-field confidence, or just the answer?

A well log is a brutal stress test, those docs are a mess of tables and scans. What happens when the model is confident but wrong on a single cell? Is there a per-field confidence flag or do you lean on the reasoning pass?

Max effort makes sense when one wrong field is expensive, like in a well log. The lever in production is knowing which pages need it, so the deep pass runs on the hard sections and the rest stays cheap and fast.

How are you scoring 'most accurate' on docs like that well log: field-level exact match, or something weighted to the numbers that matter? On long reports one wrong unit in a single table row costs more than ten formatting misses.

Spending more tokens only where the page is hard makes sense. I'd still want a rough cost per page before sending a big batch.

architectural drawings?

Max-effort extraction is still a reading, not a posting. In SAP a document number appears only after company code, document type, and a named authorization object all pass, and the change log is written in the same step. 61% accuracy is a queue, not a closed ledger.

Well logs are a brutal benchmark: dense tables, unit conversions, scanned pages and handwritten notes all in one file. The thing that matters most once this runs at scale is calibrated per-field confidence. If you are pushing millions of extracted values into a warehouse, average accuracy is less useful than knowing which 2% of cells need a human look. Curious whether Max Effort exposes that, and whether the extra reasoning tokens can be targeted only at the low-confidence regions to keep cost predictable.

Spending dynamic tokens per visual detail is the right instinct - extraction errors compound into every downstream decision. My read: as agents handle longer documents, verifiable accuracy per page becomes the actual product, not the model wrapper.

Claiming it's the "most accurate on the planet" ignores the inherent cost vs. accuracy tradeoff shown in the chart. At what point does the marginal gain in accuracy justify the extreme effort?

a well log is a brutal test, dense tables plus scanned plots. does max effort flag where it was unsure? on reports that long the useful thing downstream is knowing which pages need a human look, not only a higher average score

are you still doing " verify" i can't see anywhere

on the well log report, is accuracy scored per field or per page, and who labeled the ground truth?

Can a field return conflicting candidates with page-level provenance, so a reviewer can distinguish source disagreement from extraction error?

Can it read a prescription?

Answer keys are the slow part of a benchmark like this. A typesetter can produce them: I build one, Postext, and it returns every line and figure with its coordinates. Same values, different columns, floats and footnotes: you'd see exactly which layout breaks extraction.

Spending tokens dynamically per page makes cost hard to predict. Do users get an estimate before a big run, or a cap they can set per document?

Wow! Incredible work 👏🏻👏🏻👏🏻🫡

Spending more tokens reasoning didn't catch errors for us. Building a consensus from multiple reader did better.

"Max Effort" on every page is the opposite of cheap RAG - accuracy wins, token bill follows. Would you flip this on for contracts, or only when a human already flagged the doc?

when a multi-page table splits across page breaks, summing column totals against the footer catches dropped rows before any retry

What does a max effort run cost on a few hundred pages? Well logs are brutal to parse, half the tables are scanned at an angle and the units change halfway through the report

curious how you score accuracy on something like a 300 page well log. per field checks against a labeled set, or spot checks?

Extraction accuracy is half of it for invoices. The other half is matching what you pulled against what's already in the system, like vendor, amount and dates, so a person only sees the real mismatches. That's where most of my build time has gone.

does page order affect the extraction? a correction tucked into an appendix could change a value that looked settled near the start.