Proof · NicNacParse
We Measured Our Parser Against the Paid One. Then We Published the Score.
We ran our parser and the paid one on the same seven documents, with every test written before either system ran: NicNacParse passed 21 of 21 checks, LlamaParse 20 of 21. Then we published the whole score — including the three places the paid service still beats us.
The result: NicNacParse passed 21 of 21 mechanical checks across seven document classes. LlamaParse — a well-regarded paid service, at 10 credits a page — passed 20 of 21. Six classes tied. One class, sheet music, we won outright. And on three unscored axes — image description, heavy-scan speed, wrapped-cell tables — the paid service is still better, and we say so on the scorecard itself.
How was the test run fairly?
- The tests came first. Every acceptance check was written down before either system touched the documents. Nobody moved a goalpost after seeing a result — including us.
- One fixed scorer. A mechanical checker program graded both outputs against the sources. Along the way it caught three bugs in itself; our verdict, verbatim: the scorer was wrong more often than either parser.
- Real paperwork. A book excerpt, a church flyer, a multi-column layout, a worksheet, a sixteen-page report — plus a scanned set and a hymn score generated for the study so reading error could be measured against exact ground truth. Forty pages in all.
- Our failures are in the log. NicNacParse took four recorded passes to get here; one change made scans worse and was reverted, not buried. It's all on the published scorecard.
Where did each system win?
| Document class | LlamaParse | NicNacParse | Verdict |
|---|---|---|---|
| Book / long-form prose | 5/5 | 5/5 | Parity |
| Church flyer | 3/3 | 3/3 | Parity |
| Multi-column layout | 4/4 | 4/4 | Parity |
| Form / worksheet | 2/2 | 2/2 | Parity |
| Scanned pages (OCR error, lower wins) | 0.69% | 0.87% | Parity — both inside the pass band |
| Sheet music | 2/3 | 3/3 | Ahead — we also caught the small-print fingerings theirs dropped |
| Long report (16 pages) | 3/3 | 3/3 | Parity |
Where is the paid one still better? In print, from us.
It describes pictures. LlamaParse writes what an image shows; NicNacParse marks that an image is there, with its size. Accounted for — not described. Our closing plan (a describing pass on our own machines) is on the record.
It reads heavy scans faster. On scanned pages, accuracy ties — but our careful three-pass reading takes about 68 seconds a page; theirs takes about 2.4. For a shoebox of scanned records, that difference is real, and we print it.
Its tables are prettier. When a cell wraps to two lines, theirs merges it back into one clean cell. Ours sometimes splits it — all the words, in the right order, but the table wants a human glance. Recovered and readable; not pretty. We wrote that verdict before we wrote this page.
And the honest other direction: on ordinary text documents, ours was faster on every class — 0.2 to 1.1 seconds a page against 1.2 to 3.5 for the upload-and-parse round trip.
Why build our own at all?
Because we don't rent the tools our clients' businesses run on. Every page sent through a metered parser is rent paid on your own paperwork — and the document leaves your walls to be read. NicNacParse reads on our machines, at $0 a page, and the document never travels. The study that proved it cost 400 of their credits, total. That was tuition. We don't pay it anymore.
Don't take our word for it. Watch it read yours.
The live demo reads your PDF in front of you — pages, layout, and tables rebuilt into clean markdown. Same pipeline we run for clients, not a recording. Nothing is stored.
Try the live demo — NicNacParseProposed pricing, labeled as such: $25 per document · Studio $149/month · Practice $349/month — confirmed at intake before anything is billed. The full scorecard, methodology, and gap-closing plans: nicnacsystems.com/proof/parse-results.
Questions readers ask us
Can I try it on my own document?
Yes — the live demo on the NicNacParse page reads a PDF in front of you (about a 2 MB cap for the demo; nothing is stored). It is the same pipeline we run for clients.
What does it cost?
Proposed: $25 per document, Studio $149/month, Practice $349/month — proposed pricing, confirmed at intake. Our internal cost is $0 per page; the service we measured charges 10 credits a page at the tier tested.
Where is LlamaParse still better?
Image descriptions, heavy-scan speed (about 2.4s vs our 68s a page), and wrapped-cell table merging. All three are published on our proof page with closing plans you can hold us to.
Parse notes, when there's something measured to say
New document classes, re-run scores, and gap closures as they're proven — only when the numbers exist.
Double opt-in — a confirmation email completes your signup. Unsubscribe any time; our suppression list is global and permanent.