Yes I know and the current search syntax is pretty cool.
But the LLM search with a RAG framework would do more semantic search.
Don’t know for sure it would work but for example for my recent question on MDA/DA, I had searched for the term but not found relevant info, I should have searched for CDFA
I wonder if a embeddings/semantic search ala LLM would have worked better. In practice in other field it often does as it maps concept semantically impressively well, even if words and keywords are not the same.
For those who didn’t know I highly recommend this video to explain visually:
Practically speaking EuroGA can’t be changed cost effectively because it’s written in Ruby and anyone doing Ruby is on 1k+ per day.
BTW someone suggested moving the whole thing to another platform but that would just move us to another ongoing cost stream while abandoning a platform which has proved to be extremely reliable as well as attack resistant over 12 years.
But one could host a dedicated piece of searching code on the EuroGA Airports server (PHP).
Just watching the video. Very interesting and amazing that approach works as it does!
Peter wrote:
Practically speaking EuroGA can’t be changed cost effectively because it’s written in Ruby and anyone doing Ruby is on 1k+ per day.
I’m sure you are right, but that only shows the sorry state of the software industry. Any decently trained developer should be able to program in any reasonably mainstream language after a quick familiarisation. And Ruby is definitely mainstream. The problem is that most developers are not decently trained. (By “decent” I mean among other things that they have been taught programming language and software development principles and not simply taught to code in a single language or two.)
Cost would definitely be an issue and I don’t think it would scale well for a free public forum, though for the sake of argument:
looking at the semantic search and thread summarisation/questioning (call it thread chat)
For both you’d need a server that can run python and install packages (forget about re-implementing the llm frameworks…)
That server would need access to querying the database for the text of the post either directly or via rest or other api.
thread chat definitely the easiest, easy python to write, you’d just need to have an implementation of an LLM that can query the full text of a thread. Then you can run it locally with an open source model or send prompt to openAi or other provider which can be expensive and potential privacy issue.
semantic search would require installing a vector database, plus a rag framework, more involved, provider may do some of that for $…
And running these definitely expensive to scale it with hardware cost to run the model.
I think it would be interesting to do a bit of research whether it would even be valuable, that probably would be low cost and would be happy to contribute time. I am just interested how it would work in practice on aviation for personal intellectual curiosity.
Unfortunately reality is way more complicated, unless one is just writing the sieve for finding primesAny decently trained developer should be able to program in any reasonably mainstream language
I read every new post on this thread. I don’t know why as it’s all well above my understanding. Perhaps I’m hoping to learn/understand something. It’s not working.
I just don’t understand why we want to change this forum, IMO I think it works really well. In the past whether it’s websites or forums I have noticed that the more improvements are made for those who understand the less dinosaurs like me understand how to use the thing. KISS.
This week I took a look the Flyer Forum and IMHO I find it very clumsy by comparison and I haven’t even signed up.
Yeah some old NGO forums do look like hell.
Peter wrote :
How do the websites which detect ChatGPT etc do it? They are very reliable.
One idea would be to run the main models on the post (including extracting most probables words from the model, not just picking one), comparing the most probable next words with the actual next word. I’m pretty sure in a human post you almost always end up with a very improbable next word, or at least a very different probability signature.
I have no experience with this technique though.
I just don’t understand why we want to change this forum, IMO I think it works really well.
+1 for that. I’m not really a dinosaur, I’ve been doing computer stuff for a living for… a very long time (suffice it to say I remember paper tape and punched cards). And the longer I work with them, the more I hate them. And I still like the forum the way it is.
gallois wrote:
I just don’t understand why we want to change this forum
I don’t think the suggestion is to change this forum. At least, not the point of my own posts.
I think the point is some of us think that this technology has potential to change the way we interact with technology and information going forward. Some don’t and think will be incremental at best. Could go either way, time will tell.
But for those of us who do believe our interaction with technology is likely to change as a result of these models, it’s interesting to think how the tools, website, technology we use could be adapted or augmented, without loosing functionality as well.
Today to implement some of the potential interaction vision, the technology remains too slow and expensive and with some shortcoming. But in all likelihood this will change over time.
Peter wrote:
Unfortunately reality is way more complicated, unless one is just writing the sieve for finding primes
No. Reality is very simple. Just as I wrote most people working as software developers have not been properly trained. The reason is that demand is high and you really don’t need much training to be an acceptable coder in a single language.
