PageIndex reaches 35,621 stars arguing against vector search

PageIndex reached 35,621 stars on GitHub, built on the argument that retrieval-augmented generation doesn't need vector search at all. Version 0.2.16 released September 10 under an MIT license, with 3,141 forks and 104 open issues on the repo.

Most RAG systems chop a document into chunks, embed each one, and pull the closest matches by vector similarity when a question comes in. PageIndex skips that step. It builds a hierarchical index of a document's structure, section by section, and hands the model that outline to reason over directly.

The same day, Milvus sat at 46,067 stars and turbovec at 16,963, both pure vector-search projects still adding stars of their own. Neither camp is losing ground. People are starring the no-embedding approach and the embedding-heavy databases at once, which says more about where RAG still struggles than either project's pitch does.

Adoption numbers describe how many people starred a repo, not whether hierarchical reasoning beats vector search on messy real-world documents. That comparison isn't in the numbers here.