PageIndex

Most RAG systems break at the retrieval step. They embed everything into vectors and fetch the “closest” matches, but similarity is a poor proxy for what you actually need. PageIndex takes a different approach: it retrieves by reasoning.

Instead of vector similarity, PageIndex analyses what you’re asking for and reasons about which documents contain relevant information. This matters most for long, professional documents where the right answer might be buried in context that doesn’t match your query’s surface features.

We’ve seen this problem everywhere. Financial reports, research papers, legal texts. Traditional RAG finds passages that sound similar to your question, not passages that actually answer it. PageIndex tries to bridge that gap by making retrieval a reasoning task rather than a matching task.

It’s still experimental, but the core insight is solid. Retrieval should understand what you need, not just what you said.