Classify material by its characteristics
PDFs with selectable text, scanned pages, complex tables and illustrated manuals present different challenges. Inspect pages for skew, blur, omissions and obscured content to distinguish problems in the original from parsing errors.
Accurate OCR text alone does not establish that the page has been understood. Relationships between legends and prose, headers and values, and footnotes and conclusions can all change the answer.
Build a test set from representative pages
| Page type | What to check closely |
|---|---|
| Continuous prose | Reading order, paragraphs and heading hierarchy |
| Table | Row and column relationships, merged cells and units |
| Drawing or flowchart | Legends, labels, direction and related explanations |
| Scanned page | Text completeness, page number and recognition errors |
| Multiple manual versions | Version, effective date and applicable equipment |
Test parsing with questions
- For each page type, choose a factual question whose answer a reviewer can locate precisely.
- Compare the retrieved passage with the original and check whether it includes enough context.
- Check the page number, illustration and version cited in the answer.
- Add counterexamples with similar wording but different meanings to see whether the system selects the wrong material.
- Save error categories and representative samples. Retest after adjusting the processing method.
Route unclear content to human review
If numbers are blurred, labels are missing or versions conflict, find a clearer original first. Do not infer critical parameters from insufficient material. Check answers about operating procedures against a verified original manual and retain its scope and the need for appropriate human judgment.
Choose the parsing method, resources and parameters supported by your deployment. Some files cannot be parsed in full.
