Conversation

Notices

1. Embed this notice
   snacks (snacks@netzsphaere.xyz)'s status on Sunday, 01-Feb-2026 23:44:16 JST snacks

   in reply to

   • meso
   @meso retrieval augmented generation. You use vectorization to categorize parts of your text and can then draw up the closest matches and feed just those to an llm
   • Embed this notice
     meso (meso@new.asbestos.cafe)'s status on Sunday, 01-Feb-2026 23:44:17 JST meso

     in reply to
     @snacks rag?
   • Embed this notice
     meso (meso@new.asbestos.cafe)'s status on Sunday, 01-Feb-2026 23:44:19 JST meso
     @snacks is there an AI model you can feed large documents to and ask it questions about them. like i wanna be able to ask questions about the exact nature of the traffic laws here because it's very hard to read through that shit
     snacks repeated this.
   • Embed this notice
     snacks (snacks@netzsphaere.xyz)'s status on Sunday, 01-Feb-2026 23:44:19 JST snacks

     in reply to

     • meso
     @meso you'll prob need some kind of rag setup of you want to query over your entire traffic law tbh. Maybe there's some rag in a box thing but i'm not aware of any
   • Embed this notice
     meso (meso@new.asbestos.cafe)'s status on Sunday, 01-Feb-2026 23:48:02 JST meso

     in reply to
     @snacks you couldnt?
   • Embed this notice
     snacks (snacks@netzsphaere.xyz)'s status on Sunday, 01-Feb-2026 23:48:02 JST snacks

     in reply to

     • meso
     @meso it's production code at my company lmao
   • Embed this notice
     snacks (snacks@netzsphaere.xyz)'s status on Sunday, 01-Feb-2026 23:48:04 JST snacks

     in reply to

     • meso
     @meso if i coupd i'd give you the rag tool i made for my finals
   • Embed this notice
     snacks (snacks@netzsphaere.xyz)'s status on Sunday, 01-Feb-2026 23:52:31 JST snacks

     in reply to

     • meso
     @meso vectorization is performed by a seperate specialised ai model
   • Embed this notice
     roko's basilisk (vii@dsmc.space)'s status on Sunday, 01-Feb-2026 23:53:22 JST roko's basilisk

     in reply to

     • meso
     @snacks @meso it might be more than you need but https://github.com/HKUDS/RAG-Anything a place to start digging
   • Embed this notice
     snacks (snacks@netzsphaere.xyz)'s status on Sunday, 01-Feb-2026 23:55:34 JST snacks

     in reply to

     • meso
     @meso then you run your query through the same embedding model and get the closest matches in your db, combining bith embeddings and text search usually gives the best results, i think pgvector even has a good example how to combine them
   • Embed this notice
     snacks (snacks@netzsphaere.xyz)'s status on Sunday, 01-Feb-2026 23:55:36 JST snacks

     in reply to

     • meso
     @meso it's not that hard to implement yourself tbh, most of my time was spent wrestling file formats and shitty microsoft webservers.
     Just figure out a way to cut your documents into small enough chunks with as much meaning as possibke in tact, run it through an embedding model and save the result into a database that can handle querying vectors with like 1000 dimensions
   • Embed this notice
     protoss (nigger@detroitriotcity.com)'s status on Monday, 02-Feb-2026 00:03:04 JST protoss

     in reply to

     • meso
     @snacks @meso the lion doesn't respect IP or NDAs
     snacks likes this.
   • Embed this notice
     𝅙𝅙𝅙𝅙𝅙𝅙𝅙𝅙 (sally@freesoftwareextremist.com)'s status on Monday, 02-Feb-2026 00:29:55 JST 𝅙𝅙𝅙𝅙𝅙𝅙𝅙𝅙

     in reply to

     • meso
     @meso @snacks
     
     > like i wanna be able to ask questions about the exact nature of the traffic laws here because it's very hard to read through that shit
     
     Just learn to read.
     snacks likes this.