33
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 14 Jul 2024
33 points (100.0% liked)
TechTakes
1490 readers
30 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 2 years ago
MODERATORS
Current flavor AI is certainly getting demystified a lot among enterprise people. Let's dip our toes into using an LLM to make our hoard of internal documents more accessible, it's supposed to actually be good at that, right? is slowly giving way to "What do you mean RAG is basically LLM flavored elasticsearch only more annoying and less documented? And why is all the tooling so bad?"
Our BI team is trying to implement some RAG via Microsoft Fabrics and Azure AI search because we need that for whatever reason, and they've burned through almost 10k for the first half of the running month already, either because it's just super expensive or because it's so terribly documented that they can't get it to work and have to try again and again. Normal costs are somewhere around 2k for the whole month for traffic + servers + database and I haven't got the foggiest what's even going on there.
But someone from the C suite apparently wrote them a blank check because it's AI ...
Confucius, the Buddha, and Lao Tzu gather around a newly-opened barrel of vinegar.
Confucius tastes the vinegar and perceives bitterness.
The Buddha tastes the vinegar and perceives sourness.
Lao Tzu tastes the vinegar and perceives sweetness, and he says, "Fellas, I don't know what this is but it sure as fuck isn't vinegar. How much did you pay for it?"
The fuck's a rag in an AI context
NSFW (including funny example, don't worry)
RAG is "Retrieval-Augmented Generation". It's a prompt-engineering technique where we run the prompt through a database query before giving it to the model as context. The results of the query are also included in the context.In a certain simple and obvious sense, RAG has been part of search for a very long time, and the current innovation is merely using it alongside a hard prompt to a model.
My favorite example of RAG is Generative Agents. The idea is that the RAG query is sent to a database containing personalities, appointments, tasks, hopes, desires, etc. Concretely, here's a synthetic trace of a RAG chat with Batman, who I like using as a test character because he is relatively two-dimensional. We ask a question, our RAG harness adds three relevant lines from a personality database, and the model generates a response.
It's the technique of running a primary search against some other system, then feeding an LLM the top ~25 or so documents and asking it for the specific answer.
So you run a normal query but then run the results through an enshittifier to make sure nothing useful is actually returned to the user.
Basically
so, uh, you remember AskJeeves?
(alternative answer: the third buzzword in a row that’s supposed to make LLMs good, after multimodal and multiagent systems absolutely failed to do anything of note)
I always saw it more as LMGTFYaaS.
Maybe hot take, but I actually feel like the world doesn't need strictly speaking more documentation tooling at all, LLM / RAG or otherwise.
Companies probably actually need to curate down their documents so that simpler thinks work, then it doesn't cost ever increasing infrastructure to overcome the problems that previous investment actually literally caused.
Definitely, but the current narrative is that you don't need to do any of that, as long as you add three spoonfulls of AI into the mix you'll be as good as.
Then you find out what you actually signed up for is to do all the manual preparation of building an on-premise search engine to query unstructured data, and you still might end up with a tool that's only slightly better than trying to grep a bunch of pdfs at the same time.