File ProcessingπŸ”—

When uploading a file, it is only saved to the database together with the given meta information as a "job". The file is not visible in the ui yet. Then joex takes the next such job and starts processing it. When processing finished, the item and its files will show up in the ui.

If an error occurs during processing, the item will be created anyways, so you can see it. Depending on the error, some information may not be available.

Processing files may require some resources, like memory and cpu. Many things can be configured in the config file to adapt it to the machine it is running on.

Important is the setting docspell.joex.scheduler.pool-size which defines how many tasks can run in parallel on the machine running joex. For machines that are not very strong, choosing a 1 is recommended.


DuplicateCheck ->
Extract Archives ->
Conversion to PDF ->
Text Extraction ->
Generate Previews ->
Text Analysis

These steps are executed sequentially. There are many config options available for each step.

External CommandsπŸ”—

External programs are all configured the same way. You can change the command (add, remove options etc) in the config file. As an example, here is the wkhtmltopdf command that is used to convert html files to pdf:

docspell.joex.convert {
  wkhtmlpdf {
    command = {
      program = "wkhtmltopdf"
      args = [
        "--load-error-handling", "ignore",
        "--load-media-error-handling", "ignore",
      timeout = "2 minutes"
    working-dir = ${java.io.tmpdir}"/docspell-convert"

Strings in {{…}} are replaced by docspell with the appropriate values at runtime. However, based on your use case you can just set constant values or add other options. This might be necessary when there are different version installed where changes in the command line are required. As you see for wkhtmltopdf the page size is fixed to DIN A4. Other commands are configured like this as well.

For the default values, please see the configuration page.

Duplicate CheckπŸ”—

If specified, the uploaded file is checked via a sha256 hash, if it has been uploaded before. If so, it is removed from the set of uploaded files. You can define this with the upload metadata.

If this results in an empty set, the processing ends.

Extract ArchivesπŸ”—

If a file is a zip or eml (e-mail) file, it is extracted and its entries are added to the file set. The original (archive) file is kept in the database, but removed from further processing.

Conversion to PDFπŸ”—

All files are converted to a PDF file. How this is done depends on the file type. External programs are required, which must be installed on the machine running joex. The config file allows to specify the exact commands used.

See the section docspell.joex.convert in the config file.

The following config options apply to the conversion as a whole:

docspell.joex.convert {
  converted-filename-part = "converted"
  max-image-size = ${docspell.joex.extraction.ocr.max-image-size}

The first setting defines a suffix that is appended to the original file name to name the converted file. You can set an empty string to keep the same filename as the original. The extension is always changed to .pdf, of course.

The second option defines a limit for reading images. Some images may be small as a file but uncompressed very large. To avoid allocating too much memory, there is a limit. It defaults to 14mp.


Html files are converted with the external tool wkhtmltopdf. It produces quite nice results by using the webkit rendering engine. So the resulting PDF looks just like in a browser.


Images are converted using tesseract.

This might be interesting, if you want to try a different language that is not available in docspell's settings yet. Tesseract also adds the extracted text as a separate layer to the PDF.

For images, tesseract is configured to create a text and a pdf file.


Plaintext files are treated as markdown. You can modify the results by providing some custom css.

The resulting HTML files are then converted to PDF via wkhtmltopdf as described above.


To convert office files, Libreoffice is required and used via the command line tool unoconv.

To improve performance, it is recommended to start a libreoffice listener by running unoconv -l in a separate process.


PDFs can be converted into PDFs, which may sound silly at first. But PDFs come in many different flavors and may not contain a separate text layer, making it impossible to "copy & paste" text in them. So you can optionally use the tool ocrmypdf to create a PDF/A type PDF file containing a text layer with the extracted text.

It is recommended to install ocrympdf, but it also is optional. If it is enabled but fails, the error is not fatal and the processing will continue using the original pdf for extracting text. You can also disable it to remove the errors from the processing logs.

The --skip-text option is necessary to not fail on "text" pdfs (where ocr is not necessary). In this case, the pdf will be converted to PDF/A.

Text ExtractionπŸ”—

Text extraction also depends on the file type. Some tools from the convert section are used here, too.

Text is tried to extract from the original file. If that can't be done or results in an error, the converted file is tried next.


Html files are not used directly, but the converted PDF file is used to extract the text. This makes sure that the text is extracted you actually see. The conversion is done anyways and the resulting PDF already has a text layer.


For images, tesseract is used again. In most cases this step is not executed, because the text has already been extracted in the conversion step. But if the conversion would have failed for some reason, tesseract is called here (with different options).


This is obviously trivial :)


MS Office files are processed using a library without any external tool. It uses apache poi which is well known for these tasks.

A rich text file (.rtf) is procssed by Java "natively" (using their standard library).

OpenDocument files are proecessed using the ODS/ODT/ODF parser from tika.


PDF files are first checked for a text layer. If this returns some text that is greater than the configured minimum length, it is used. Otherwise, OCR is started for the whole pdf file page by page.

docspell.joex {
  extraction {
    pdf {
      min-text-len = 500

After OCR both texts are compared and the longer is used. Since PDFs can contain text and images, it might be safer to always do OCR, but this is something to choose by the user.

PDF ocr is comprised of multiple steps. At first only the first page-range pages are extracted to avoid too long running tasks (someone submit an ebook for example). But you can disable this limit by setting a -1. After all, text that is not extracted, won't be indexed either and is therefore not searchable. It depends on your machine/setup.

Another limit is max-image-size which defines the size of an image in pixel (width * height) where processing is skipped.

Then ghostscript is used to extract single pages into image files and unpaper is used to optimize the images for ocr. Unpaper is optional, if it is not found, it is skipped, which may be a compromise on slow machines.

docspell.joex {
  extraction {
    ocr {
      max-image-size = 14000000
      page-range {
        begin = 10
      ghostscript {
        command {
          program = "gs"
          args = [ "-dNOPAUSE"
                 , "-dBATCH"
                 , "-dSAFER"
                 , "-sDEVICE=tiffscaled8"
                 , "-sOutputFile={{outfile}}"
                 , "{{infile}}"
          timeout = "5 minutes"
        working-dir = ${java.io.tmpdir}"/docspell-extraction"
      unpaper {
        command {
          program = "unpaper"
          args = [ "{{infile}}", "{{outfile}}" ]
          timeout = "5 minutes"
      tesseract {
        command {
          program = "tesseract"
          args = ["{{file}}"
                 , "stdout"
                 , "-l"
                 , "{{lang}}"
          timeout = "5 minutes"

Generating PreviewsπŸ”—

Previews are generated from the converted PDF of every file. The first page of each file is converted into an image file. The config file allows to specify a dpi which is used to render the pdf page. The default is set to 32dpi, which results roughly in a 200x300px image. For comparison, a standard A4 is usually rendered at 96dpi, which results in a 790x1100px image.

docspell.joex {
  extraction {
    preview {
      dpi = 32
Please note

When this is changed, you must re-generate all preview images. Check the api for this, there is an endpoint to regenerate all preview images for a collective. The cli tool can be used for this.

Text AnalysisπŸ”—

Finally, the extracted text is analysed to find possible metadata that can be attached to the new item. There are two different approaches provided.

The basic idea here is, that instead of you defining textual rules to apply tags and other things, these rules are found for you based on what you have provided so far.

Docspell relies on the Stanford NLP Library for its AI features. Among other things they provide a classifier and NER annotators. The latter is also a classifier, that associates a label to terms in a text. It finds out whether some term is probably an organization, a person etc. It tries to β€œunderstand” the structure of the text, like verb, nouns and their relation.

The two approaches used are sketched below. They have both advantages and disadvantages and are by default used together. However, depending on the document languages, not all approaches are possible. They also have different memory footprints, and you might want to disable some features if running on low-end machines.


If you enabled classification in the config file, a model is trained periodically from a collective's files. Very roughly speaking… this model contains the essence of "patterns" in the text that are likeley related to a tag, a corrpesondent etc.

When a new document arrives, this model is used to ask for what metadata (tag, correspondent, etc) it thinks is likely to apply here.

Training the model is a rather resource intensive process. How much memory is needed, depends on the number of documents to learn from and the size of text to consider. Both can be limited in the config file. The default values might require a heap of 1.4G if you have many and large documents. The maximum text length is set to 5000 characters. If all your documents would be that large, adjusting these values might be necessary. A model is trained periodically, the schedule can be defined in your collective settings. Although learning is resource intensive, using an existing model is quite cheap and fast.

For tags, you can define the tag categories that should be trained (or that should not be trained). Docspell assigns one tag (or none) from all tags in a category to a new document.

Note that tags that can not be derived from the text only, should probably be excluded from learning. For example, if you tag all your items with Done at some point, it may falsly learn patterns to this tag and tag your new documents with Done.

The admin can also limit the number of documents to train with in the config file to control the memory footprint when training.

Classification is used in Docspell once for guessing tags and also for finding correspondent and concerned entities. For correspondent and concerned entities, the NLP approach is used first and the classifier results then fill missing values.

Natural Language ProcessingπŸ”—

NLP is the other approach that works a bit differently. In this approach, algorithms are used that find lanugage properties from the given text, for example which terms are nouns, organization or person names etc. This also requires a statistical model, but this time for a whole language. These are also provided by Stanford NLP, but not for all languages. So whether this can be used depends on the document language. Models exist for German, English, French and Spanish currently.

Then Stanford NLP also allows to run custom rules against a text. This can be used as a fallback for terms where the statistical model didn't succeed. But it can also be used by itself. Docspell derives these rules from your address book, so it can find terms in the document text that match your organization and person names. This does not depend on the document language.

By default, Docspell does both: it first uses the statistical language model (if available for the given language) and then runs the address-book derived rules as a last step on so far unclassified terms. This allows for the best results. If more than one candidate is found, the "most likely" one is set on the item and others are stored as suggestions.

The statistical model approach works generally very well, i.e. for large address books. Normally, a document contains only very few organizations or person names. So it is more efficient to check these few against your (probably large) address book; in contrast to testing hundreds of company names against a single document. It can also find things not in your address book (but this is unused in Docspell currently). However, it might not detect all or there are no statistical models for your language. Then the address book is used to automatically create rules that are run against the document.

Both ways require memory, it depends on the size of your address book and on the size of the language models (they vary for each language). In the config file, you can specify different modes of operation for nlp processing as follows:

  • mode full: creates the complete nlp pipeline, requiring the most amount of memory, providing the best results. I'd recommend to run joex with a heap size of a least 1.4G (for English only, it can be lower that that).
  • mode basic: it only loads the NER tagger. This doesn't work as well as the complete pipeline, because some NLP steps are simply skipped. But it gives quite good results already and uses less memory. I'd recommend to run joex with at least 500m heap in this mode.
  • mode regexonly: this doesn't load any statistical models and is therefore much lighter on memory (depending on the address book size, of course). It will use the address book to create regex rules and match them against your document. Memory usage then doesn't depend on the document language.
  • mode disabled: this disables nlp processing. Then only the classifier is run (unless disabled).

Note that mode full and basic is only relevant for the languages where models are available. For all other languages, it is effectively the same as regexonly.

The config file allows to specify a limit for texts to analyse in general. Large texts result in higher memory consumption. By default, the first 10'000 characters are taken into account.

Then, for the regexonly mode, you can restrict the number of address book entries that are used to create the rule set via regex-ner.max-entries. This may be useful to reduce memory footprint.

The setting clear-stanford-nlp-interval allows to define an idle time after which the language models are cleared from memory. This allows memory to be reclaimed by the OS. The timer starts after the last file has been processed. If you can afford it, it is recommended to disable it by setting it to 0.