A table can survive PDF extraction with every number intact and still become misleading. If the column heading falls off a page or gets attached to the wrong row, a search result may return a plausible value with the wrong unit or reporting period.
This walkthrough uses KolmoPDF's PDF-to-Markdown converter and a small audit fixture to check that a table remains usable for retrieval. The fixture is illustrative, not a claim about a particular conversion result.
Pick a page that can fail
Choose a document you have permission to upload. Find a table with a unit in its heading, a footnote, or rows that continue onto another page. Note its printed page number and PDF page number: front matter can make them different. Keep the original PDF so a reviewer can compare each result with the source.
For example, suppose the source has this table, with the final row on the following page:
| System | Latency (ms) | Error rate (%) |
|---|---|---|
| A | 120 | 1.8 |
| B | 95 | 2.1 |
| C | 110 | 0.9 |
An extracted row such as B | 95 | 2.1 is incomplete on its own. The values need the column names and units. If the source also qualifies those measurements in a footnote, the qualification belongs in the evidence a reader can inspect.
Convert and inspect
Open the converter, upload the PDF and convert it to Markdown. The service handles technical layouts including tables and formulas. Download the result and locate the table by its caption or a distinctive row label. Check the rendered Markdown and the raw text: a pipe character inside a cell, for example, can change how a Markdown table renders unless it is escaped.
Compare the output against the source:
- Does each value sit under the right heading and unit?
- Do the row labels and values continue in the correct order across a page break?
- Are merged cells, footnotes and qualifiers preserved or clearly represented?
- Can you trace the table back to the original pages?
A continuation page may produce a second table. If it lacks headings, add the original headings to that fragment before indexing, but keep a pointer to the pages from which its rows came. Do not assume two nearby tables are one continuation: compare captions, column structure and surrounding text first. When the source is ambiguous, send it for review instead of joining rows automatically.
The converter processes documents online. Use an approved workflow for confidential material, and check the output on representative pages before processing a larger collection. Complex layouts and scans deserve particular attention; conversion does not guarantee that every cell or formula will be correct.
Build retrieval chunks around evidence
A chunk containing only B | 95 | 2.1 asks the retrieval system to infer too much. A more useful chunk repeats the table title, heading and relevant note:
Source: report.pdf, PDF page 8
Table: Service measurements
System | Latency (ms) | Error rate (%)
B | 95 | 2.1
Note: Values are from the illustrative fixture.
Keep provenance in metadata as well as in text if your retrieval system supports it. Use the actual PDF page range for each chunk, including both pages when a row or note crosses a boundary. If the Markdown export does not supply reliable page references, capture them during review rather than assigning a guessed page.
Repeat headings when splitting a long table into row groups. Keep a footnote with every group it qualifies, and retain the original full table for inspection. For a table with multi-level headings, flatten the heading path explicitly, for example Latency / p95 (ms), after checking that it describes the original cell.
Test before indexing the collection
Write a few questions with answers you can verify in the PDF. For the fixture above, ask for system B's latency and its unit, then ask which system has an error rate of 0.9%. Check the retrieved chunk before looking at the generated answer. It should contain the correct row, the heading and the source reference. Also ask a question for a value absent from the table; the system should have an opportunity to say the evidence is missing rather than filling in a number.
Record the failures by type: extraction, table reconstruction, chunking, retrieval or answer generation. A correct answer with the wrong citation still needs work. Re-run the same questions after a parser or chunking change, and include examples from scanned pages and multi-page tables if those appear in your documents.
The KolmoPDF product overview covers the available output types. For a separate check on mixed scanned and text pages, see page-level OCR routing.
Top comments (1)
The B / 95 / 2.1 example also needs an explicit relationship between each footnote and the cells it qualifies. Repeating every note with every row can introduce a different extraction error when a note applies only to one marked column or a subset of systems.
I would add two otherwise identical rows to the fixture, with the footnote marker on only one of them. Ask the same measurement question for each row and inspect which qualifier reaches the answer. That catches overly broad note propagation even when the numbers, headings, units, and page references all survive.