OrthoScience – Hybrid search engine for 500K+ orthopedic translational research
orthoarchives.com
orthoarchives.com
Existing search engines rely on keyword matching or citation-based ranking. Neither approach handles cross-disciplinary queries well.
A search for "PEEK versus titanium spinal fusion cage mechanical properties" returns thousands of irrelevant clinical results or misses the engineering paper that matters.
So I built OrthoScience.
Query routing, not just retrieval. The pipeline classifies each query before executing it. DOI patterns, quoted titles, and author prefixes route to exact-match lookups and never touch the semantic layer. Free-text queries hit a hybrid pipeline: full-text (tsvector) and vector similarity (pgvector) run in parallel, results fuse via reciprocal rank fusion, then pass through a multi-signal adaptive ranking function.
Why adaptive? Broad queries (thousands of results) produced near-identical relevance scores, making the top-10 essentially random. The ranking function now detects query breadth and shifts weight toward quality signals for broad queries, relevance signals for specific queries. Score discrimination improved 4.7x for broad queries with zero regression on specific ones.
The translational problem. Clinical and translational queries behave differently at every layer. "FEA" rarely appears in abstracts because authors write "finite element analysis" in full. Domain-specific synonym expansion solved recall (+32% for technical queries, zero clinical regression). Evidence hierarchies differ too: an in-vitro biomechanical study is gold standard in translational context, not a limitation. The ranking function adjusts evidence weights by search mode. A cross-encoder reranker helped clinical queries (+4.3% nDCG) but hurt translational ones (-7%), so reranking activates conditionally.
Evaluated on a human-labeled gold set: nDCG@10 = 0.955 across clinical, material science, simulation, and abbreviation query categories.
It's free to use and requires no sign-in.