How we read 1,401 pages of scanned manuals and matched 99.3% of part numbers exactly

The right part depends on the serial number.

Field technicians ask for a part number on a phone and expect the right one in under half a minute.

The client sells an assistant to equipment dealers and rental fleets. Technicians ask it for part numbers, specifications and procedures. Many of the manuals are old scans with wrong or missing text stored inside the file.

The right part depends on the machine model, serial range and fitted options. Two rows that look identical can apply to different machines. Fixing the answer to one question could also break another answer a few months later.

Every page read from the picture, twice over.

Our first decision was to ignore the hidden text inside the files completely. We turned every page into an image and read it the way a person would.

The client chose parts catalogues and a maintenance manual from its own library, 1,401 pages in all, including a black and white scan whose hidden text was gibberish.

Two versions of Claude read every page separately, with a different maker’s model used for a third check. Each reading recorded the parts rows, applicable machines and serial ranges, and cross-references in the same format. The third check covered the 490 riskiest pages, including every page of the worst scan.

Our checking code compared the readings field by field, and every disagreement went back to the page image to be settled. We added the third reader because two readers agreeing can still both be wrong. It paid off: it caught pages where both had made the same slip, usually a single character, like an F read as a P.

Checked against the manuals’ own index.

Most of the manuals print their own part number index at the back. We used those indexes as a reference for checking the extracted part numbers.

Of 5,571 part numbers in those indexes, 5,530 appeared in our database exactly, character for character. That is 99.3% of the indexed part numbers.

The database holds 14,537 parts rows and 11,099 part and model numbers. It also includes a plain-English summary of each manual and a list of the machines each manual covers.

The technician can check the answer on the source page.

The assistant asks for missing machine details before choosing a part.

The assistant gives one of four responses: an answer, a question, a refusal when it cannot answer, or a reference to the right place. Every answer cites the manual, printed page, page in the file and exact row. The technician checks the source and decides which part to use. For harder questions, the assistant checks the page image before replying.

Give it a serial number and the manuals that don’t apply fall away. Ask about a pump that comes in two voltages and it asks which one, instead of guessing.

We hold every answer to a strict standard: the right response, every part number exact, the right manual and the right page. In a live session, the client’s founder put his own question to it, and the answer matched the one he already had.