Conversion options and performance

Docling’s defaults favor quality. When you process many documents, or know something about your inputs, a few options trade quality for speed – or buy you images and higher resolution.

The Python backend

doclingr wraps the Docling Python library through reticulate. Install the backend once into a managed environment, then restart R:

library(doclingr)

install_docling()         # creates an "r-docling" virtualenv
# ...restart R...
docling_available()       # TRUE

Model cache and offline use

The deep-learning models (layout, tables, OCR) download on first conversion – several hundred megabytes – and are cached afterwards, so only the first run pays the cost. The cache lives in the standard Hugging Face location (~/.cache/huggingface by default). To control where models are stored, set HF_HOME before the first conversion (and before Python initializes):

Sys.setenv(HF_HOME = "~/.cache/doclingr-models")

On locked-down or air-gapped machines, populate the cache once from a machine with access (run one conversion of each kind you need, or copy the HF_HOME directory over), then force offline mode so nothing tries to reach the network:

Sys.setenv(
  HF_HOME        = "/shared/doclingr-models",
  HF_HUB_OFFLINE = "1"
)

With HF_HUB_OFFLINE = "1", conversions use only what is already in the cache and fail fast (instead of hanging on a blocked connection) if a model is missing. Both variables must be set before the first conversion in the R session; the safest place is .Renviron or the shell environment.

OCR

OCR reads text from scanned pages and images. It is on by default. For born-digital PDFs (exported from Word, LaTeX, etc.) the text layer is already present, so turning OCR off is a large, safe speed-up:

doc <- docling_convert("born-digital.pdf", ocr = FALSE)

Leave OCR on for scans, photographs of documents, or anything where text is “painted” into an image.

Table structure: accurate vs. fast

# Best structure (default) -- complex, spanning, nested tables
docling_convert("report.pdf", table_mode = "accurate")

# Quicker -- clean grids, large batches
docling_convert("report.pdf", table_mode = "fast")

Hardware acceleration

Pick the device the models run on, and optionally the CPU thread count:

docling_convert("report.pdf", device = "mps")               # Apple Silicon
docling_convert("report.pdf", device = "cuda")              # NVIDIA GPU
docling_convert("report.pdf", device = "cpu", num_threads = 8)

device = "auto" (the default) lets Docling choose.

Images and figures

By default images are not retained, which keeps results small. Ask for them when you want to save figures or work with page images:

doc  <- docling_convert("paper.pdf", images = TRUE, images_scale = 2)
figs <- docling_figures(doc, image_dir = "figures")
figs

images_scale = 2 renders at roughly twice 72 DPI; raise it for crisper figure exports at the cost of memory.

Batch conversion

Pass a vector of sources to convert them in one batch; the result is a named list of documents:

docs <- docling_convert(
  c("a.pdf", "b.docx", "c.html"),
  ocr        = FALSE,
  table_mode = "fast"
)

length(docs)
docs[["a.pdf"]]

A pragmatic recipe

For a large pile of born-digital reports where you mostly care about text and tables:

docs <- docling_convert(
  list.files("reports", pattern = "[.]pdf$", full.names = TRUE),
  ocr        = FALSE,      # no scans
  table_mode = "fast",     # clean grids
  device     = "auto"
)

Then chunk and embed as shown in vignette("rag").