Project to Automatically Select Tags for Zettelkasten Documents

Motivation for the Project

When working within a zettelkasten, it’s not uncommon to neglect the task of tagging documents. This can be attributed to the speed of creating a new card, leaving little time to tag it properly. Additionally, it’s easy to overlook certain tags that may be relevant at that moment. There’s also the consideration of whether it would be beneficial to create new tags. So what if I implement a feature to automatically tag documents within a zettelkasten using ChatGPT?

However, there are potential downsides to this approach. For instance, in some people’s implementations, certain tags might be reserved, creating a potential for confusion. Additionally, the AI might misinterpret the context, leading to inaccuracies.

I also wanted to implement this as a feature within the zkvr project. I have a couple ideas for AI powered scripts that could run within an interconnected graph of Zettelkasten documents.

Results

Moving onto the results, the system can occasionally surprise you with its creativity by adding tags that don’t yet exist. The time it takes to run for each document is relatively short, only a second. However, it’s not always accurate, at times adding tags that are only loosely related to the content. Sometimes it returns duplicate tags, or returns some of the same tags that already exist. Despite these minor drawbacks, it maintains a relatively good level of accuracy overall.

Code

Generating the Prompt for ChatGPT

First, I created a simple script for generating the prompt for ChatGPT. It simply needs to provide:

• a list of tags that the assistant can select from
• some text around the prompt to provide the instructions
• the document
• a constraint on the number of tags returned.
  • if this is not constrained, it might return 2 or it might return 20. it would be annoying to clean up if it added too many.

I could do some further prompt engineering here if I start having problems with the tags it returns.

#!/bin/bash

echo "Using the following set of hashtags,"
cat $(root-knowledge-base-repo-path)/all_tags
echo 'choose an appropriate set of at most five tags for the following markdown document:'
echo ''

cat "$1"

Generate Tags for Document Passed as Argument

I used the gpt CLI command tool that I selected for use in the terminal

#!/bin/bash

# get current script directory
SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )

# use a grep pattern to pull out tags to keep output format consistent
$SCRIPT_DIR/generate-prompt "$1" | gpt --model gpt-3.5-turbo --prompt - | grep -Eo '#[A-Za-z]+' | sed 's/^#//g'

That last line might look scary, so let’s dive into it a bit.

Scraping the Resulting Tags from the Output

GPT might return the tags in any format. Sometimes, there is one tag per line, and sometimes, it puts all the tags together. Sometimes there’s a preamble that says “here are the appropriate tags” or something similar, and sometimes there is not. Therefore, it is necessary to scrape the resulting tags out.

• scraping out tags

... | grep -Eo '#[A-Za-z]+' | sed 's/^#//g'

The one consistent part of the GPT output is that it always has hashtags. The first part of this line returns all tags in the document, one per line. The second removes the leading hashtag from each line.

• example GPT output before scraping

Based on the content of the markdown document, the appropriate tags would be:

- `#script`
- `#chatGPT`
- `#automation`
- `#markdown`
- `#tags`

Implementing it in ZKVR

I added a new command to the program called autotag. It will automatically add tags to the current document. I was able to reuse a script that I created for removing duplicate tags during the process of implementing the merging functionality

It might be a good idea to add some additional configuration steps to zkvr to setup a ChatGPT API key so the cool features work if someone else uses zkvr.

• example of programmatically adding tags to a card by ID from scripts using zkvr’s CLI tools

./zc addtag -t test 20231114070621

• adding it to zkvr user interface code

    temptags=$(mktemp)
    echo "Generating tags..."
    # generate tags using GPT and store in temp file
    zet/20231114070621/generate-tags "$doc" > "$temptags"
    echo "Tags:"
    cat "$temptags"
    # add tags from GPT output
    cat "$temptags" | xargs -IR -n 1 ./zc addtag -t R "$id"
    # remove duplicate tags, reuse tempfile
    zet/20231121064457/fix-duplicate-tags "zet/$id/README.md" > $temptags
    # overwrite original document
    mv $temptags "zet/$id/README.md"