[GUI Plugin] AI Assisted Tags

[p0lrb3ar]

I have an obsessive-compulsive book tagging problem! If a book is in my library, I need to have some way to search for it in my library server, aside from the title and author, which I often forget. Keyword searching should be the answer, but how to get them to be consistent and mostly-accurate became the struggle, and that struggle was real!

Metadata sources all use different methodologies to classify books. They all have their own keyword systems, book categories, subject headings, etc... It's all nonsense if you ask me, but my delimma remains. While I'm lounging beside the empty pool at my sketchy Tiajuana motel, I want to search for books about training goldfish in Naval warfare tactics! I know I have a book about that somewhere!

This plugin can help you get there by using a little assistance from AI. This plugin uses whatever Provider and Model you have configured in the Calibre AI Provider plugins. I personally use my local LM Studio with Qwen3-Coder-Next or Qwen3-30b-a3b-2507 or Gemma-4-e2b. Results WILL vary!

First, what AI is not. Magic or even the slightest bit helpful without decent prompting.

AI suffers from:

• Hallucinations (Fabrication of Information)
• BIAS and Hubris (included in it's stolen training data)
• Lack of True Understanding or Reasoning
• Outdated Knowledge
• Inconsistency and Non-Determination
• Prompt Sensitivity
• Resource Starvation
• Safely and Political Alignment Issues
• Context Limits
• Anthropomorphism
• And probably a lot MORE.

Taking these issues into account, and not being able to get my AI the professional help it needs, the plugin needed to be grounded in something real. This turned out to be the BISAC Subject Heading List from the BISG (NFW RLY? YR!) which I included as a textfile based on their publicly available data. No trademarks were treaded on in the making of this plugin.

It uses two AI passes to assign the mostly likely subject headings to a book based on Title, Author, and Comments. Testing against numerous models averages between 6-10 seconds per book. More details on how it works are in the ABOUT.txt file.

Even this is not enough sometimes. You may get a tag that says, "NON-Classifiable". If you see this, run the plugin again, then again, and possibly again, and again, then, change your model to something different because it's experiencing one of the issues above, or just imaganing a world where all humans are in battery storage containers. Seriously, AI dislikes us. But who knows, AI just sucks most of the time and has terrible logging. That's why you have to demean... uh prompt it.

There are the plugin configurations.

• Max tags per book (defaults to 5)
• Tag behavior: Replace or Append
• [optional] Skip books with no description
• [optional] Debug logging in calibre-debug.exe.
• Requests per minute (0=Unlimited)
• Tokens per minute (0=Unlimited)
• Test the AI Provder, sends a hello to the configured provider/model and waits for a reply.

Notes on AI Configuration:
This plugin uses the Calibre AI Provider plugins. Which ever provider and model you use (free or paid), setting the temperature or reasoning effort to a lower setting like Temperature: .1 or .2, Reasoning Effort: Low or No Reasoning makes the model more deterministic (factual). Higher temperatures make a model more creative and diverse which you might want for your latest FanFic epic, but it is not something you want when choosing the best option from a list of options, like this plugin does.

The plugin runs as a background job, so you can stop it when it is acting up. You can also assign a keyboard shortcut to it. It really has all the bells and whistles. Please enjoy!

Updates:

• 7/2/2026 - Added Rate limiting and Token limiting to assist with using free models (e.g. OpenRouter, Ollama, Cerebras, etc...)
• 7/2/2026 - Added additional detection for bad JSON responses to reduce the number of Non-Classibiable results (LM Studio)

[p0lrb3ar]

Reserved