Time in a LLM reasoning engine

AI Discussions
,
1

Does anyone knows how good could be LLM in writing summaries, by connecting some text events with time?

Example:
Assume, I have a huge amount of logs in OpenSearch to train LLM with short summary from operators. At some point, I want to load some piece of logs (e.g Warnings only) and get a short summary:

  1. “Service has problems with auth for 10 min” or “Service is warming up now”
  2. Or even just classification “service is Healthy/Unhealthy since 14:23:00”.

It turns out, order and timing in such classification are really important. Do you know any models that could be good with a such kind of tasks?

Question is similar to /llm-for-timeseries-data/713248, but here we talking about mix of time and text.

Edit 1: I can, of course upload logs in batches and write a summary as an aggregation of those batches: e.g batch #5 that have logs since 14:23:00 classified as “unhealthy”, after four “healthy” batches and I make a conclusion that service is not OK since 14:23:00. But I would love to have it done for me automatically, if possible.

2

Did you experiment with Chatgpt?

3

Yes, I finally did it. I’ve just finished my small investigation of this topic in ChatGPT.
Chat GPT (v4.0) do an amaizing job in writing summary of what is going on in a logs in general(for example, identify if there was a hacking attack attempt), but fail with more precise timing and order

I did a little bit of prompt engineering, asking model to return to me:

  • time range when event did happened
  • how long did it last.
  • did some events happened simultaneously, in parallel.
  • did one event happen before other event (e.g restart had happened before a server failure)

It did fail in most of these task. It could group events by date, but not by time.

4

Indeed, it can happen. Perhaps if you check our free courses in this site you may find some app that can further help, I haven’t checked them myself to be frank.