#pest | 2025-11-16

00:31

Not a terrible summary: https://paste.rs/cVk10.md. Seems to be confused about a tangentially related debugging discussion.

~ 11 hours 12 minutes ~

11:44

signpost

might need to chew the logs into embeddings to get good RAG over them. maybe the MCP server already does this?

11:52

awt_akris

signpost: don't know what you mean be embeddings.

11:52

awt_akris

@!ask In the context of LLMs and RAG what are embeddings?

12:04

raip

awt_akris: https://paste.rs/8XG6H.md

12:04

signpost

typical search poopline would involve processing the messages using an embedding model, obtaining a vector which is a stab at representing the text's position in a vector space learned by the embedding model, storing that embedding alongside the text

12:05

signpost

then when you want to query, having an LLM elaborate the user's query a bit (to produce more text that has a chance of being similar to embeddings in your db), and then using e.g. cosine distance to rank database messages by similarity, taking the top however many, then feeding those to an LLM to synthesize a response

~ 2 hours 11 minutes ~

14:16

*

signpost yaps further to kick off another backlog sync

~ 36 minutes ~

14:53

signpost

(implemented prod on boot so I don't have to keep doing that)

~ 4 hours 11 minutes ~

19:04

awt_akris

Ok started generating embeddings. Responses seem a bit more thorough.