AI Toolcast releases a free checker for one invoice extraction exercise

A fixed exercise, from source to report

AI Toolcast has released AI Extraction Checkup, a free browser exercise built around one synthetic invoice. Readers download the source, give it and an explicit instruction to an AI service they already use, then paste the returned JSON into the checker. No signup or API key is required by the checker.

The purpose is to make a small extraction result inspectable. Instead of assigning a general model ranking, the page shows which expected values are missing or different and which outputs break the requested representation. AI Toolcast owns and develops this tool; this article is an announcement of our own product.

Source 1Source 2

The instruction is part of the test

The fictional four-row invoice contains identifiers with leading zeroes, a wrapped description, quantities, a discount, tax and shipping. Its checked total is $147.47. The instruction asks for JSON strings, dates in YYYY-MM-DD form and two decimal places for monetary values.

One explicit convention matters: discount_amount is the positive magnitude to subtract, even though the invoice displays a minus sign. The checker labels the opposite sign as a representation-rule finding rather than presenting it as evidence of a generally inaccurate model. Required values that cannot be read should be returned as null; the report marks that uncertainty instead of treating a guessed replacement as success.

Source 1Source 2

What the comparison preserves

The rules cover 37 scalar fields. Rows are matched by their unique printed line_id, so reversing row order does not create an error. Documented descriptive whitespace is normalized, while identifiers retain meaningful characters and leading zeroes. Decimal values are compared through strings rather than rounded binary floating-point numbers.

Duplicate keys or conflicting row identifiers do not get silently resolved into a preferred answer. Missing rows, extra rows, changed values and format or type deviations receive separate findings. A syntactically accepted response can therefore produce a useful differences report without being a match. Even a complete match says only that this particular response met this particular exercise.

Source 1Source 2

Prepared controls make the rules visible

Six labeled controls demonstrate the behavior: a matching answer, harmless whitespace and row reordering, a missing last row, a changed identifier, swapped quantities and a wrong total. The changed-identifier control, for example, produces one value finding and 36 matching fields out of 37. These are deliberately constructed examples, not recorded failures from a named AI model.

The open source includes the comparator, original fixture and generation source, instruction, schema and tests. The release was checked with 37 comparison tests, 15 event-collector tests and browser checks covering the supplied controls, malformed input, literal rendering, keyboard use and report exports. Those checks test the implementation; they do not establish reader demand or performance across real invoices.

Source 1Source 2

Local results, limited measurement

The pasted response and generated report stay in the page unless the reader chooses to copy or download them. There are no customer-document uploads, provider calls, saved response history or paid features in this release. The AI service used for the extraction has its own data policy and may have its own fees or access limits.

On the hosted page, optional usage events record fixed action names and allowlisted source or campaign categories with a random identifier held in the open tab. They exclude pasted text, report contents and field values. Built-in examples and marked QA are excluded from reader completions. Global Privacy Control and Do Not Track disable this collector, and the exercise still works if collection fails.

The next decision is whether people find the complete exercise useful. We will review recorded task activity and distribution before expanding it. Custom documents, hosted extraction, broader evaluation and paid extensions are outside this pilot; source availability is not evidence that those features or a commercial offer exist.

Source 1Source 2