Last updated 4 April 2021


The main daiR vignettes use deliberately simple examples involving uploads of pdf files straight into the root of the bucket and down again. In real life you may be dealing with slightly more complex scenarios.

Image files

Document AI accepts only PDFs, GIFs and TIFFs, but sometimes your source documents are in other formats. daiR’s helper function image_to_pdf() is designed to help with this. Based as it is on imagemagick, it converts almost any image file format to pdf. You can also pass a vector of image files and ask for a single .pdf output, which is useful for collating several pagewise images to a single, multipage .pdf.

To illustrate, we can take this image of an old text from the National Park Service Website:

download.file("https://www.nps.gov/articles/images/dec-of-sentiments-loc-copy.jpg", 
              destfile = "nps.jpeg", 
              mode = "wb")

And convert it to a pdf like so:

library(daiR)
image_to_pdf("nps.jpeg", "nps.pdf")

And the file is ready for processing with Document AI.

Processing a folder tree

At other times you may want to have folders inside your bucket. A typical scenario is when your source documents are stored in a folder tree and you want to batch process everything without losing the original folder structure.

Problem is, it’s technically not possible to have folders in Google Storage; files in a bucket are kept side by side in a flat structure. We can, however, imitate a folder structure by adding prefixes with forward slashes to filenames. This is not complicated, but requires paying attention to filenames at the upload and download stage.

To illustrate, let’s create two folders in our working directory: folder1 and folder2:

library(fs)
dir_create("folder1")
dir_create("folder2")

Then we create three duplicates of the file nps.pdf and put two pdfs in each folder.

file_move("nps.pdf", "./folder1/nps.pdf")
file_copy("./folder1/nps.pdf", "./folder1/nps2.pdf", overwrite = TRUE)
file_copy("./folder1/nps.pdf", "./folder2/nps3.pdf", overwrite = TRUE)
file_copy("./folder1/nps.pdf", "./folder2/nps4.pdf", overwrite = TRUE)

To upload this entire structure to Google Storage, we create a vector of files in all subfolders with the parameter recurse = TRUE in the dir_ls() function. I’m assuming here that the working directory is otherwise empty of pdf files.

pdfs <- dir_ls(glob = "*.pdf", recurse = TRUE)

We then iterate the gcs_upload() function over our vector:

library(googleCloudStorageR)
library(purrr)
resp <- map(pdfs, ~ gcs_upload(.x, name = .x))

If we now check the bucket contents, we see that the files are in their respective “folders”.

Bear in mind, though, that this is an optical illusion; the files are technically still on the same level. In reality, the folder1/ and folder2/ elements are an integral part of the filenames.

We can process these files as they are with the following command:

resp <- dai_async(pdfs) 

In which case DAI returns .json files titled folder1/<job_number>/0/nps-0.json and so forth. We can download these the usual way:

content <- gcs_list_objects()
jsons <- grep("*.json$", content$name, value = TRUE)
resp <- map(jsons, ~ gcs_get_object(.x, saveToDisk = .x))

And the json files will be stored in their respective subfolders alongside the source pdfs.

Note, however, that this last script only worked because there already were folders titled folder1 and folder2 in our working directory. If there hadn’t been, R would have returned an error, because the gcs_get_object() function cannot create new folders on your hard drive.

If you wanted to download the files to another folder where there wasn’t a corresponding folder tree to “receive” them, you would have to use a workaround such as changing the forward slash in the bucket filepaths for an underscore (or something else) as follows:

resp <- map(jsons, ~ gcs_get_object(.x, saveToDisk = gsub("/", "_", .x), overwrite = TRUE))