ZK Vector DB (ignore)

This is a proof of concept for a decentralized vector database. Web2 AI applications would be wildly inefficient if reproduced on-chain; instead, we use SP1 to create proofs for the execution. SP1 made it easy to generate and verify ZK proofs with minimal understanding of the topic!

Motivation

Existing market options and drawbacks:

1. Perplexity.ai is an amazing AI-powered search engine

• TLDR: an LLM that is focused on summarizing relevant websites and cites its sources
• But it's centralized. Why is this an issue?
• Lack of privacy; web search data is often a cornerstone of demographic modeling for big tech
• Reliance on the company to maintain data quality and continually process new websites

1. Privacy-focused search engines: decentralized ones like Presearch, self-hosted ones like Searx, etc.
   • Not popular, often due to poor results quality: low relevance or not returning latest websites

The end goal for this PoC is a decentralized version of Perplexity, which simultaneously solves the problems in the prior options.

Proposal

The project is split into two components:

1. Self-hosted LLM
   • users are free to use whatever models they want to do summarization, including self-hosted open-source models like Llama
   • due to the nature of the UX, half of the results quality concerns are handled by the LLM itself (fixable in prompt engineering)
2. Decentralized vector database
   • a public DB for website results, for the users
   • assuming appropriate incentives
   • users can freely host nodes
   • users can contribute new websites to the DB, addressing the concerns with data availability

This repo contains a webapp to demonstrate the end use case with the LLM, as well as a simple vector database built in Rust.

Built With

• express.js
• javascript
• ollama
• react.js
• rust
• sp1