
Jev, PageIndex Tackle Long-Document Search
VectifyAI’s new Apache-2.0 project combines Jev’s structured decisions with PageIndex’s hierarchical document trees to locate answers in reports too large for flat prompts. The examples search NVIDIA and Citigroup filings without vector databases or embeddings.
This is a smart division of labor: Jev handles bounded relevance choices while PageIndex supplies the navigation structure, turning long-document RAG into hierarchical search.
- –Avoids flat-search limits of 32K tokens and 255-choice inputs
- –Uses section summaries and page ranges to progressively narrow the search
- –Keeps multiple candidates through beam search before validating answer-bearing pages
- –Demonstrates the approach on 93-page and 318-page financial filings
- –The concept is compelling, but the repository’s printed evaluation covers only two examples
DISCOVERED
1h ago
2026-10-02
PUBLISHED
1h ago
2026-10-02
RELEVANCE
AUTHOR
typesafeai