Source: dsRAG (D-Star AI) README

Open-source retrieval engine combining semantic sectioning, contextual chunk headers (AutoContext) and relevant segment extraction, with FinanceBench and KITE results.

Contents
  • What: open-source project README (vendor benchmark claims)
  • Author/publisher: D-Star AI (Zach McCormick, Nick McCormick)
  • Date: read 2026-11-14
  • Original: github.com/D-Star-AI/dsRAG

Summary

dsRAG targets dense documents such as financial reports, legal documents and papers. At ingest, an LLM splits the document into titled sections (semantic sectioning), and AutoContext builds a contextual chunk header for each chunk from a generated document title and summary plus section titles (and optionally section summaries); the header is prepended before embedding. At query time, Relevant Segment Extraction (RSE) stitches clusters of adjacent relevant chunks into longer contiguous segments.

Key claims

  • Contextual chunk headers are deterministic templates of document- and section-level context, not a per-chunk free-text blurb.
  • FinanceBench: 96.6% with dsRAG (Claude 3.5 Sonnet, mostly default settings) vs 19% for vanilla RAG in the original benchmark paper and 32% in the authors’ own baseline.
  • KITE average score (0-10): top-k baseline 4.72; RSE 6.73; CCH + top-k 6.04; CCH + RSE 8.42. Semantic sectioning was not used in these runs.
  • RSE contributes slightly more than CCH; the combination is best.

Feeds

Contextual retrieval

Suggest changes →