Tessera: training-free layer rebuilds retrieved HTML into evidence units for RAG
_reachsumit · x · 2026-10-29
The paper traces answer evidence from source HTML to reader input across English Wikipedia and finds RAG chunking loses evidence—especially tables and infoboxes separated from their headers—and stronger extraction, context expansion, or bigger reader budgets don't fix it. Tessera is a training-free evidence delivery layer: it rebuilds retrieved pages' HTML into fine-grained evidence units completed with page title, section path, and table headers, then selects and packs query-relevant units into the reader context at an unchanged token budget. Index, retriever, and reader stay untouched; it works for text, screenshot, and agentic pipelines, improving accuracy across three QA benchmarks.
More from coding & agent
- Pydantic's Code Mode report: 51% faster long tasks across 8 models and 5,276 runs — samuelcolvin · 2026-10-29
- Decagon details a self-improving autoresearch loop for inference stack optimization — kimberlywtan · 2026-10-29
- Dev uses Claude-built library to ship Obsidian plugin that plans music practice with spaced review — dSebastien · 2026-10-29
- Sitzmann open-sources texwerk, a self-hosted collaborative LaTeX editor with native agent support — vincesitzmann · 2026-10-29
- Professor runs DeckWerk's headless server mode for real-time lab slide collaboration — vincesitzmann · 2026-10-29
- Rime launches Prism, a speech-to-speech model that kills dead air in voice agent calls — testingcatalog · 2026-10-29