More accurate search inside papers
Papers are now split along their real structure and the top search results are re-ranked by a dedicated relevance model, so assistants land on the passage that answers the question.
When an AI assistant connected through MCP searches PaperVantage's library, it gets back passages from papers. Two changes make those passages more likely to be the right ones.
Structure-aware chunking
Papers are now split along their own headings, so a passage is a coherent piece of the paper rather than an arbitrary slice of text. Along the way:
- The reference list and other back matter are no longer indexed, so a search for a topic doesn't return a bibliography entry.
- Code blocks and table captions stay together with the text they belong to.
- Chunks are far more even in size, which keeps the best passage from being buried in a huge one.
Re-ranking the best candidates
Search first finds candidates by meaning and by keywords, then a relevance model re-orders the top thirty so the most useful passage comes first. If the re-ranker is ever unavailable or slow, search quietly keeps the order it already had instead of failing.
Re-ranking sends the search query and the public passages being ranked to a third-party model provider. Your private uploads' text is not sent. See the Privacy Policy for details.