RAGenesis
A retrieval-augmented generation platform centred on the verses of five of the most significant books in history: the Torah, the Bhagavad Gita, the Analects, the New Testament and the Quran. Users explore and compare the texts through semantic search and generation, in what I call a "generative book" compiling the wisdom of many traditions.
Beyond the app, RAGenesis proposes a framework for transparency of embedded knowledge bases. It defines Semantic Similarity Networks (SSN), a special case of semantic networks that exposes the graph implicitly created by the choice of embedding model, knowledge base and similarity metric, and derives a chunking methodology from elementary graph theory. Follow-up work measures semantic hypervolume through convex hulls of embeddings and defines intertext similarity and consistency measures between texts.
Built solo, end to end, in ten weeks: Streamlit front end, Milvus vector database, all-MiniLM-L6-v2 and jina-clip-v1 embeddings, deployed on AWS (EC2, ECR, CloudFormation/SAM) with Docker and GitHub Actions. Open-sourced under MIT.