A PDF looks correct when someone opens it. The headings are clear, the table has sensible columns, and the scan is readable. Yet the document assistant gives an answer that mixes two rows or misses a page entirely.
Before replacing the language model, inspect the material it received. In a text-based document-search pipeline, the PDF first becomes extracted text and structured records. That conversion can lose information before the search or answer-writing step begins.
We build RAG systems that include document ingestion as well as retrieval and evaluation. We recommend treating ingestion as something to test explicitly, rather than assuming that an accepted upload means a document was read correctly.
Start with the page, not the upload count
Choose a few pages whose meaning someone on the team can verify. Include an ordinary paragraph, a table, a multi-column page, and a scan if those occur in the collection. Keep the original visible beside the representation used for search.
For each page, ask a practical question with a known answer. Can the extracted material still answer it without guessing? Counting files or words will not tell you whether a number remains attached to the correct label.
Consider a hypothetical service company comparing delivery options in a supplier’s PDF. The document lists a delivery area, a service level, and a cutoff time in adjacent columns. If those columns become separate lists, every word might survive while the relationships disappear. This is an invented example, not a JTPCK client account.
Separate digital text from scanned images
Some PDFs contain usable digital text. Others contain images of printed pages, which need optical character recognition, or OCR, before a text-based search can use them. A single packet can require different treatment on different pages.
PyMuPDF’s OCR guidance warns that forcing recognition on clean digital text can slow processing and reduce output quality. We recommend checking the existing text before replacing it, and testing image-based pages rather than assuming that one extraction method suits the whole packet.
A quick copy-and-paste check can expose garbled or missing text. It is only a screening step. It does not prove that every page, image, or annotation has been included in the actual indexing pipeline.
When a page is unreadable, retain that status. An empty extraction should not quietly become evidence that the original page contains nothing useful.
Read the extracted text in order
PDF text extraction does not necessarily follow the order a person sees on the page. PyMuPDF documents cases where headers added later appear after the body in extracted output. It also exposes text positions that can help reconstruct a useful sequence.
Inspect the sequence directly. Does a sentence continue into the correct column? Does a sidebar interrupt a procedure? Does a repeated footer separate a heading from its paragraph?
We recommend correcting these problems before dividing the material into searchable passages. Otherwise, the passage boundaries may preserve a broken sequence. Making passages longer can carry more text without restoring the original relationships.
For the hypothetical delivery guide, the test is whether the extracted passage expresses one complete option at a time. A fluent summary produced later cannot establish that the underlying columns were joined correctly.
Keep table labels with their values
A table needs more than a list of recognized words. The row, column, header, unit, and nearby qualification may all determine what a value means.
Microsoft’s layout documentation describes outputs with table cell positions, header roles, page numbers, and bounding regions. These provide useful ingredients for preserving relationships. They do not remove the need to compare the result with the original.
AWS also identifies merged cells and inconsistent table layouts as sources of unreliable extraction. Include those layouts in the review set when they appear in your documents.
We recommend testing questions that depend on a row label and a column label together. Then test a value qualified by a note below the table. Microsoft’s documentation notes that table bounding regions can exclude captions and footnotes, so a highlighted table alone may leave out relevant context.
The desired result is an inspectable relationship between a value and its meaning, not merely a table-shaped output.
Preserve a route back to the original
Keep a stable document identifier and source-page reference with the extracted material. If the parser supplies coordinates, use them to help reviewers locate the passage. Make the difference between the PDF’s page position and a printed page label clear when they differ.
We recommend retaining the original separately from any cleaned or corrected text. Record which processing version produced the searchable representation. That makes it possible to compare a later change without losing the evidence used for an earlier answer.
Corrections should be visible. If a person fixes a misread character, the system should distinguish that reviewed correction from the raw extraction. Do not replace an unreadable value with a plausible one simply to make the record complete.
Give failed pages a destination
A document should not have only two states, uploaded or missing. We recommend tracking which pages were processed, which need inspection, and which failed. Reconcile those results with the expected page count.
That check will not prove that every sentence is correct. It does make omissions easier to notice. For important material, have the responsible reviewer decide whether a partial document can be searched at all and what warning users should see.
Sometimes the cheapest fix is a better scan or the original digital file. Sometimes a layout-aware parser is justified. For a small collection, manual correction may be more practical than a more elaborate pipeline.
Our AI document analysis work includes scanned and unstructured records. If your assistant struggles with PDFs, tell us which page types cause trouble. Start with a description of the failure and arrange an appropriate exchange before sharing confidential samples. The useful first question is what information was lost, and at which step.