Extract PDF tables into analysis-ready data

8 GB RAM works for ordinary papers; 16 GB is safer for long scanned reports. GPU is optional.

Component	Role	Price	Link
Docling 2.117.0	layout-aware PDF and table extraction	Free / 免费open-source local processingchecked 2026-08-02	https://github.com/docling-project/docling/releases/tag/v2.117.0
DuckDB 1.5.5	schema inspection, validation and export	Free / 免费MIT licensed in-process databasechecked 2026-08-02	https://github.com/duckdb/duckdb/releases/tag/v1.5.5

Extract and profile tables

uv venv --python 3.12 .venv
. .venv/bin/activate
uv pip install 'docling==2.117.0' 'duckdb==1.5.5'
mkdir -p extracted
docling input.pdf --to json --to md --output extracted
# After exporting each reviewed table as extracted/table_001.csv:
duckdb -c "CREATE OR REPLACE TABLE t AS SELECT * FROM read_csv_auto('extracted/table_001.csv', header=true); COPY (SUMMARIZE t) TO 'extracted/table_001_profile.csv' (HEADER, DELIMITER ','); COPY t TO 'extracted/table_001.parquet' (FORMAT parquet);"

Table extraction audit

Compare the extracted table with the named page in the source PDF. Check title, header hierarchy, row labels, column order, merged cells, footnotes, units, decimal signs, missing-value symbols and superscripts. Return a correction list with page number, row key, column key, extracted value, visible source value and confidence. Do not fill values that are unreadable.

Known pitfalls

• Visually merged cells often need manual normalization.
• A plausible CSV can still have shifted columns; validate totals and row counts.
• Keep page and table identifiers beside every export.
• Do not coerce symbols such as <0.01 or NR into numeric nulls without a documented rule.

No local accelerator required. Suitable for batch workloads with approved cloud processing.

Component	Role	Price	Link
Google Cloud Layout Parser current managed processor	managed structure and table-oriented layout extraction	Usage-based / 按用量计费$10 per 1,000 pageschecked 2026-08-02	https://cloud.google.com/products/document-ai/pricing
DuckDB 1.5.5	local validation of downloaded structured output	Free / 免费local open-source validation layerchecked 2026-08-02	https://github.com/duckdb/duckdb/releases/tag/v1.5.5

Table artifact acceptance contract

artifact_version: table-extract.v1
required_fields:
  - source_pdf_sha256
  - source_page
  - table_index
  - table_title
  - columns
  - rows
  - footnotes
checks:
  page_exists: true
  headers_nonempty: true
  unique_row_keys: preferred
  preserve_inequality_signs: true
  preserve_missing_value_tokens: true
  require_human_review_for_merged_cells: true

Known pitfalls

• Layout Parser and OCR can incur separate charges.
• Managed extraction still needs visual review for scientific tables.
• The $10 rate is per 1,000 pages, not per table.
• Store source hashes so corrected source PDFs do not silently reuse stale tables.