RelativeDB
Company publishing research on conditioning search reranking directly on relational/structured data rather than treating candidates as text. Its research post introduces RT, an 85M-parameter Relational Transformer, and releases a pretrained checkpoint (stanford-star/rt-j) as open code. Site sections observed: Research, Cloud, and a public code repository link, suggesting a hosted product alongside the research output — not confirmed beyond what the article states.
Research Contributions
- Relational Transformer (RT) — 85M-parameter reranker conditioning on typed “Cell” facts (
name, type, value) and schema relationships between database rows, rather than serializing candidates to text - LLM-based query parsing that converts free-text queries into structured Cells before reranking, reported to substantially improve RT’s accuracy over unparsed text input
Articles
- Relational Reranking - Scoring Search Results with Structured Facts — Daniel Henneberger; introduces RT, its Cell representation, ablations, and benchmark results on MTEB DeepPlanning and STaRK
People
- Daniel Henneberger — author of the RT research post
Concepts
Relational Transformer · Reranking