I'd hate to be that guy... but an LLM might be useful

[ds-sebastian]

It might be cool if unique error messages were stored and potential fixes were searched automatically via an LLM + Search so they can be reviewed later.

I understand this changes the scope of the project and would certainly impact the small footprint it has on a server.

[amir20]

It's ok, I am usually that guy.

I understand this changes the scope of the project and would certainly impact the small footprint it has on a server.

I am ok increasing scope if I think directionally it is the right choice.

I still need to understand what you mean though.

It might be cool if unique error messages were stored and potential fixes were searched automatically via an LLM + Search so they can be reviewed later.

Can you give me an example?

For example, maybe there is a log error with "something happened in code on line XX". Then dozzle has a button that automatically copies the error and searches it using your choice of LLM and spits out possible solutions?

Is that what you mean? So really the win here is copy and pasting to AI?

Or do you mean something completely different where the whole scope of the logs, image and container details are sent to an agent to further process.

There are so many options here that I think it would be good first understand the usecase.

[ds-sebastian]

I envisioned this:

• Dozzle indexes or stores unique errors (perhaps STDERRs or detected in some other manner), not including duplicates probably via some similarity scoring
• These would be stored in a different part of the UI organized by containers
• Periodically these error logs (perhaps with lines above and below) along with the container information would be sent to an LLM via API (that optionally has search tooling for more accurate diagnoses and suggestions)
• The LLM through prompting, would output explanations of the errors (possible causes, meaning, etc) and possible solutions or debugging suggestions would be stored along these tracked errors

The error tracking would be helpful in itself as a standalone feature to get a more focused look at what's going wrong, and the LLM component can be an additional addon by entering an API key and a base url for which model to use. I'm not really a developer so I often do this with logs anyway to learn what's going on and what might be potential debugging steps.

I think it would be a neat feature to have this information at a glance, because I'm usually going to dozzle to debug or finding out what's wrong anyway

[amir20]

I have been thinking about this. It would be great to find a concrete use case. For example, some times I do find myself copy/pasting logs to ChatGPT. Is the goal to automate this? What's an example of this? And what would Dozzle provide that is better.

Personally, I don't find a close integration with an LLM a highly needed feature. I do wonder if there are any other apps that do this well. Do you know of any?

[mjraneri]

I use LLM to analyze logs. A simple prompt in Google Gemini can sometimes produce very helpful results.. For instance, my typical prompt would be:

"Do you see anything unusual or alarming in these log entries?"

And then I paste 100-200 log entries. It's a good idea to check for personally identifyable information like UserIDs and things like that which you don't want fed back into the model. I'm not sure I would use a LLM that sent all the log text (not cleaned) to some central server, but a locally running LLM that, for instance once per day, analyzed logs for that day and reported any suspicious/uncommon/alarming entries could be useful. How useful depends on the LLM's depth. But I do see the value. FWIW.

[NatoBoram]

An exemple implementation of this is Gemini in Google Cloud, which can access some logs and answer questions about them.

Though, this seems like a big feature requiring multiple intertwined and not-so-small changes:

• The UI needs a new sidebar for that LLM
• There needs to be a way to select/filter logs that can be sent to the LLM
• There needs to be a way to configure which LLM to use (maybe just an OpenAI endpoint+key can be enough since everyone is copying that API)

This sounds like the kind of thing that would require a community member to step up to implement it after a design is selected.

[amir20]

This sounds like the kind of thing that would require a community member to step up to implement it after a design is selected.

I think this would be a valuable addition. I personally haven't found a lot of time to play around with Gemini API.