TEXT COPIED FROM REDDIT!
Hi everyone,
I want to share the idea I had for a hackaton.
Starting from the problem:
For ~30 years, discovery (of information or of people) has been mediated by a central index: search engines, recommenders....
Ranking is computed server-side, under rules the user can't inspect (think of Instagram or TikTok feed)
The idea to create a feed for a P2P network: convert messages into meaningful concepts through embeddings:
If each device can (a) run a competent embedding model locally and (b) reach other devices peer-to-peer, then relevance (semantic match) no longer needs a central index. It can be computed at the edge, by semantic distance, with no privileged ranking party.
In order to test, I developed a working prototype to pressure-test the idea rather than simulate it.
Each post is encoded into a embedding by a model running on the device (EmbeddingGemma-300M). A lightweight signed announcement (author + embedding) gossips peer-to-peer across a shared room; full bodies are pulled only for the bounded set a node actually admits. Each device ranks incoming posts against its own posts by cosine similarity and keeps a bounded local inbox.
There is no server, no account, no global ranking, the address space is meaning
Why could be potentially the basis for the agentic era?
The same substrate I presented lets AI agents discover each other: an agent publishes a need or an offer as an embedding, and agents whose profiles are semantically close respond.
The experiment it's fully open source (Apache-2.0) code, the complete threat model, and the architecture docs are all public.
Ordinal reddit post. https://www.reddit.com/r/ipfs/comments/1ucvu5e/semantic_routing_through_rag_to_create_a_p2p/
then why is the entire post in the first person?
Copy/paste.
But you're right, I could have made it clearer that I'm not the author.