Feature Question (AI)

Even though Hazel is a powerful tool, all rules are based on rigid matches (text, name, properties). What is missing are context-based rules, e.g. via the connection to AI tools.
Is something like this planned on the roadmap?

Can you be more specific about what you expect out of such a feature? AI is thrown around a lot without any thought to whether it's solving your problem.

You're right - maybe I'll briefly describe the use case:

I have tried to automate my entire document management as much as possible.
This involves searching through static folders that always contain the same type of document. For example, bank statements. These can be renamed and moved according to relatively simple rules. There are more rules scanning incoming documents (from emails, scanner etc)

Hazel rules can only recognize and re-name around 40 to 50% of all incoming documents.
For the rest of the documents, however, it is not worth creating separate rules, as in this form they occur rarely or only once.
For those remaining documents, I imagine that the context of the document can be evaluated in a similar way to the analysis in ChatGPT (“Briefly summarize the content of the PDF” --> “This is an invoice from company xyz for ABC”). I could then instruct Hazel to save the document based on the feedback according to my existing naming logic:

In the (simplified) example: "Invoice xyz for ABC 2024-11.pdf" and then move it using the existing rules.

Currently, I have to do the naming manually by checking each individual document manually. Afterwards my distribution rules take over again (probably I only have to store 5% of all documents myself)

Was I able to get across what I meant?

BTW: I'm aware that this can be done by scripts (e.g. AppleScript, Python, or shell scripts) already so AI models or services can be integrated into these scripts, for example by using APIs such as OpenAI or other ML tools. What I mean is a more Hazel integrated way....

Thanks for elaborating. What if ChatGPT is inconsistent with its summary or is just plain wrong? Are you willing to train the model yourself?

Mr_Noodle wrote:Thanks for elaborating. What if ChatGPT is inconsistent with its summary or is just plain wrong? Are you willing to train the model yourself?

Sure, that will happen. I would maybe start with reviewing the filename before finally sorting. Not sure if I would train the LLM myself. Maybe it would be best to start with simple use cases. For all basic documents chatGPT returned overall satisfying results.

I think there are meaningful use cases which are far away from simple AI slogans.

I could think of general support in the application, like error detection and optimization, where the AI checks existing rules for conflicts, redundancies or inefficient processes and suggests improvements. Hazel could also analyze files and suggest rules based on user behavior or file patterns that can be adopted or adapted with a single click.

But the bigger value might be pre-configured options to connect to existing tools (GPT /Gemini etc) - the token to access the service is just configurable by the user.

I would see several information valuable for AI analysis:

1. new "AI" Date fields like creation date or due date (e.g. on invoices)
2. Sender / Creator of the file
3. Category: e.g. “invoice”, “image”, “contract”, ideally from a pre-defined list (so if I have 20 categories defined and the document cannot matched to one of them it will be named "unknown", so I have to look at it)

Im pretty sure there are much more possibilities.
The advantage would be that the current logic of Hazel could be retained but Hazel would become much more powerful.

Since I love Hazel, I just wanted to start the discussion. I think it would be great if Hazel could continue to develop and become even better known...

I think AI in its current state is better suited to smaller, more directed tasks. Extraction of various data fields is definitely an option. There are also issues with licensing and whether processing is done on device or uploaded to a server. Having the user plug in their AI of choice is an interesting idea though not sure if there's any sort of common API for that (haven't looked at it from that angle so don't really know). Right now, though, if you can script it, that would be the way to integrate it.