PageIndex is a vectorless retrieval approach for documents with reliable native structure. It builds a hierarchy from sections, subsections, and page ranges, then lets an LLM agent reason over that tree to select relevant content. section 10.7.3
Its retrieval workflow exposes focused tools to the agent:
- get_document() supplies document metadata. section 10.7.3
- get_document_structure() reveals the hierarchical tree. section 10.7.3
- get_page_content(pages) retrieves only the required page ranges. section 10.7.3
Where it fits
Apply PageIndex to research papers, manuals, and contracts when their structural boundaries are clean. Avoid relying on it for unstructured prose, where little inherent hierarchy exists to guide the agent. section 10.7.3
By removing embeddings, clustering, and generated summaries, PageIndex avoids summary drift and embedding-pipeline complexity. Its tradeoff is dependence on clean document structure and agent reasoning at query time. section 10.7.3
