Extracting tables from documents
In doclingr: Document Intelligence via 'Docling'

knitr::opts_chunk$set(collapse = TRUE, comment = "#>", eval = FALSE)

Tables are where document intelligence earns its keep: they carry the numbers, but they are exactly what naive text extraction mangles. doclingr uses Docling's table-structure model to recover cells, then hands each table back as a tibble.

The basics

docling_tables() returns a list with one tibble per detected table, in document order:

library(doclingr)

doc <- docling_convert("financials.pdf")
tables <- docling_tables(doc)

length(tables)      # how many tables Docling found
tables[[1]]         # the first table, as a tibble

Each tibble carries a page attribute recording where the table came from:

attr(tables[[1]], "page")

Accurate vs. fast table structure

The table model has two modes. The default "accurate" recovers complex structure (spanning cells, nested headers) at some cost; "fast" is quicker and often enough for clean grids:

doc_fast <- docling_convert("financials.pdf", table_mode = "fast")
docling_tables(doc_fast)[[1]]

Working with the extracted tables

Because each table is a tibble, the whole tidyverse is available. For example, tag every table with its page and stack them into one long frame:

library(dplyr)
library(purrr)

all_tables <- docling_tables(doc) |>
  imap(\(tbl, i) mutate(tbl, .table = i, .page = attr(tbl, "page"))) |>
  list_rbind()

all_tables

Or write each table to its own CSV:

tables <- docling_tables(doc)
iwalk(tables, \(tbl, i) readr::write_csv(tbl, sprintf("table-%02d.csv", i)))

Tips

Column types come back as character; coerce with readr::type_convert() or dplyr::mutate(across(...)) once you know each table's schema.
If a scanned (image-only) PDF returns empty tables, make sure OCR is on (docling_convert(..., ocr = TRUE), the default).
For very wide tables split across chunks during RAG, docling_chunk() can repeat the header row -- see vignette("rag").