RAGTREC BioGenPyseriniLuceneLocal LLMsCitation ValidationNext.js

BioEvidence AI: Local-First Biomedical Question Answering

A full-stack biomedical QA application combining retrieval over a 14.9M-record PubMed index, adaptive reranking, local LLM generation, and citation validation.

2026-08-011 min read

Overview

BioEvidence is the productized, local-first continuation of my TREC BioGen work. It turns biomedical questions into evidence-linked answers while keeping model inference local and the retrieval trail visible. Its Examples mode also provides a detailed view of the TREC system's questions, retrieved evidence, answers, and citations.

System design

  • Pyserini/Lucene retrieval over a 14.9M-record PubMed index
  • Adaptive reranking and evidence selection
  • Local LLM generation with on-device inference
  • Citation validation for generated answers
  • Live integrations with PubMed/NCBI, Europe PMC, and ClinicalTrials.gov
  • Detailed TREC BioGen examples for inspecting the full question-to-citation path

What was built

  • Real-time query-to-citation pipeline spanning PubMed, Europe PMC, and ClinicalTrials.gov
  • User-facing Next.js application for live answers and saved TREC example inspection
  • Local-first inference path designed for privacy and control

Result

The TREC research pipeline and its evaluation artifacts now connect directly to a practical AI application with inspectable evidence and direct links to biomedical sources.