Dewey improves AI research by preserving document structure instead of flattening PDFs into disconnected chunks

Standard RAG approaches fail for complex research tasks because they discard document hierarchy (sections, subsections) and treat PDFs as bags of paragraphs

Dewey treats documents, sections, and chunks as first-class API primitives with section manifests that include full heading hierarchies and byte offsets

Agents can now skim document structure cheaply before committing to full chunk retrieval, enabling multi-hop reasoning across papers and synthesis of conflicting results

The /research endpoint runs an agentic loop that can traverse entire corpora at 'exhaustive' depth for citation-backed answers grounded in original source passages