← Back to work
Independent project · 2025

Scholar Assistant

Hybrid vector and keyword retrieval, with an interface that shows its sources by default.

  • Gemini 2.5 Flash
  • ChromaDB
  • Hybrid retrieval
  • FastAPI
  • React
Problem

A knowledge assistant that answers confidently but can't show its work isn't trustworthy enough to actually use for anything that matters, and pure vector search alone tends to miss exact-match queries, names, IDs, specific terms, that keyword search catches easily.

The interaction decision

Show sources by default, not behind a sources toggle a user has to think to click. If the answer can't be traced to something retrieved, that's a signal the interface should surface, not smooth over.

Architecture

A multi-turn conversational assistant on Gemini 2.5 Flash, with hybrid retrieval combining ChromaDB vector search and keyword search so semantic and exact-match queries both land, and iterative prompt evaluation tightening grounding and consistency across a corpus of religious scriptures. Python and FastAPI backend; React and TypeScript frontend that renders citations inline with the answer.

What shipped

A conversational assistant where every answer arrives with its sources attached, built to be checked rather than trusted blindly.