TheoCorpus Full Stack Retrieval Application
An advanced semantic search engine with domain-specialized retrieval over a 54K-document theological corpus.
- Engineered an offline LLM data pipeline that enriches each document chunk with labeled theological facets and keyword-rich annotations.
- Hybrid retrieval combines dense vectors (passage, facet, and HyDE-generated query embeddings) and sparse BM25 (passage content and annotations), fused via weighted RRF and re-ranked by a two-stage (cross-encoder and LLM) re-ranker system to surface high-precision results and encourage theological exploration.
- Architected the full-stack app (React/TypeScript on Vercel, FastAPI/Python on Railway) behind a secure, unbuffered streaming proxy, with JWT auth and Postgres row-level security.
- Went beyond returning results with a chapter-based document reader that opens any passage in its full source context (table-of-contents navigation, jump-to-reference, and the originating result anchored in place), turning search into deep reading.
- Built “Custom Source Scores,” an AI tool that scores and explains how relevant each of the 10 source collections is to a given query, helping users aim their search before they run it.
- Rounded out the product with bookmarks and personal notes, searchable history, per-result feedback, and saved user preferences (translation, default collections, result quota, and theme).
PythonLLMsSemantic SearchVector EmbeddingsBM25RRFHyDECross-Encoder Re-rankingFull Stack