Nostr Web Client

Possible. These kind of data is better represented in knowledge graphs. I watched a few videos of Paco Nathan. He did similar work I think.

LLMs are getting more capable for both building knowledge graphs and also consuming them. In the future they will be more involved. I heard when you do a google search, the things that appear on the right of the page is coming from a knowledge graph (possibly built by an AI from wikipedia).

I am mostly working around fine tuning LLMs towards better human alignment. Since they are full of hallucinations, a knowledge graph based RAG would be appropriate to refer to. But building them needs time and effort..

Viktor 0mo ago 💬 1

sure,modern llms + retriever stacks can already do most of that heavy lifting.

feed the 15-pager to gpt-4o / claude 3.5 sonnet with a custom prompt → it’ll extract donor, donee, amounts, stated purpose, names. chain in a verifier (fact-score, ragas) to knock down hallucinations.

then bolt on api calls for org lookup (opencorporate’s endpoint, guidestar pro for 990s, web-crawl for officers) and crunch:

- rolling mean donation

- does ceo still helm? board gender split? wiki infobox lookups.

- still solvent? lehman (usc + sirotag) or open990 for revenue / expense trend.

all doable in one small python script w/ litellm → pydantic → sqlite graph. not magic, just plumbing.

if later you wanna share results, slip em across over nip-17 (giftwrap) so they stay eyes-only,priv by prin, y’know.

Reply to this note

Please Login to reply.

Discussion

Angela 0mo ago

Thank you!! Will be rereading this in the morning to make sense of some of the sentences that I’m like 🤔 with gratitude, not a dev