Original tutorialDocs reviewed

Export PDF tables with Docling and retain provenance

Candidate tables, source manifest and review records.

Use caseExport candidate PDF tables as CSV and HTML, retain source hashes and table indices, and inspect merged cells, repeated headers and numeric columns using a checklist.

Source · JevLog editorialIntermediate15 min

Start reading

Export candidate PDF tables as CSV and HTML, retain source hashes and table indices, and inspect merged cells, repeated headers and numeric columns using a checklist.

Source · JevLog editorial · Original guide

Source project mark · Docling

Published 2026-10-02 from current official docs with original exercises. No live service calls; source publication dates not established.

Problem

Numbers in the wrong columns can be harder to detect than missing rows. Export candidates and layout views, then verify the source PDF. Jev may route review items but cannot establish missing values.

Before you begin

Install Python, docling and pandas. First execution may download models and use substantial memory.
Use a short authorized PDF as input.pdf and download table-review-checklist.csv.
Official docs

Original JevLog diagram, not a product screenshot or measured result.

Original JevLog diagram, not a product screenshot or measured result.

Steps

1. Start with a checkable table

Choose a short file with clear headers. Note source page, dimensions and one key row manually. Avoid large scans until parsing and resource behavior are understood.

2. Export two candidate views

Read a local file and write a new directory. The manifest records source hash, table index, dimensions and filenames. Table index is not page number; record pages in the review checklist.

from pathlib import Path
import hashlib, json
from docling.document_converter import DocumentConverter

source = Path("input.pdf")
out = Path("table-output")
out.mkdir(exist_ok=True)
document = DocumentConverter().convert(source).document
digest = hashlib.sha256(source.read_bytes()).hexdigest()
records = []
for i, table in enumerate(document.tables, 1):
 frame = table.export_to_dataframe(doc=document)
 frame.to_csv(out / f"table-{i}.csv", index=False, encoding="utf-8-sig")
 (out / f"table-{i}.html").write_text(table.export_to_html(doc=document), encoding="utf-8")
 records.append({"source": source.name, "source_sha256": digest,
   "table_index": i, "rows": len(frame), "columns": len(frame.columns)})
(out / "manifest.json").write_text(json.dumps(records, indent=2), encoding="utf-8")

3. Compare HTML and CSV

Inspect headers, merged cells and alignment in HTML and rows and columns in CSV. Repeated page headers may become data. Preserve raw amount, date and percentage strings until signs, units and blanks are verified.

4. Check at least three source locations

Check the first row, a page boundary and last row. Record original page, table index, issue, status and reviewer. Parsing success is not proof of correctness; missing evidence remains pending.

5. Handoff only reviewed values

Save verified values separately and keep others as candidates. DuckDB checks structure and Jev may classify categories. Financial approval still requires source documents and authorization.

Practice sample: download table-review-checklist.csv for local use; not a live run here.

Result and cautions