That last state in the State Function is where Lambda gets a nicely formatted text. It's either from OCR or in markdown from a text PDFs (extracted by pymupdf4llm package).
The OpenAIs API does not have anything to index and manage documents (I suspect they dump all files into S3 analog on Azure).
It's ok, I want a custom UI anyways, to let users manage docs and a backend to catalog them.
UI is pretty easy to build.
Now the hard part is a searchable catalog of documents. I want something like what MacOS has but not form-fields based search, just tags that makes sense to a user
I want a pile of docs from where I can choose tags like "maintenance docs", "boiler"
Or "proposals", "companyName"
Or "email", "airline", "cityName"
And then backend will attach a bunch of docs to file_search and gets ready for chat
Maybe full text search on docs short summaries.
I keep asking myself: maybe RAG-it-all and search document by context instead of metadata?
Because a versatile RAG is notoriously difficult to build. And pretty expensive to run because of server-based tech behind it (all kinds of databases)
So the alternative is a smart backend attaching a bunch different file types to a file_search. Which automatically get indexed and becomes available in chat is _really compelling_
It requires a little bit more upfront cataloging, which needed in any case for a basic security
(who's the docs owner, re-ingesting, updates, etc)
All these "home-made" RAG solutions btw pretend security doesn't exist. BigCloud takes a good care of security, but falls short on flexibility and -lessness.
And the damn cost is too high, even to play with it.
To sump up. With a doc catalog, unproven, expensive solution of building a full scale RAG pipeline became a known, cheap "index and search tons of S3 files" problem.
What's the best architecture there? I see AWS put a lot of Athena-based solutions.
@threadreaderapp unroll
• • •
Missing some Tweet in this thread? You can try to
force a refresh
A huge advantage of OpenAI is a fully Serverless RAG. There are no databases or storage to manage. File uploaded to projects (max 100GB for the whole org, can be adjusted). Then attached to vector stores entities, wait till indexed, and voila - it can do RAG
I clicked "Introducing Amazon Q Apps (Preview)" and now I can not unsee it
It sends us to a YouTube video where a team struggles with daily tasks but finds relief in Gen AI product. First, they get onboarding plan for John at Example Corp - looks great!
Example Corp was a marketing corporation but it's just a shell; in reality, it's a premier financial services company! Wut?
It partners with AnyCompany to implement AI Boost solution (hey, kids - this is what consulting looks like).
They made an Investment Analyst Assistant.
Then, another team member uses AI magic to create a sales script. It's a really powerful one: can be used for AnyProduct.
AnyProduct is a cutting edge generative AI tool, obviously. ExampleCorp wants to buy it.