A new approach to AI document search is getting attention from developers. It’s called Vectorless RAG, and it works very differently from older methods. Traditional RAG systems break documents into chunks and use math-based similarity scores to find answers. Vectorless RAG skips all of that. Instead, it uses an AI to reason through a document’s structure like a tree.
The system builds a tree index from a document. Each part of the tree has a title, page range, summary, and a node ID. When someone asks a question, the AI looks at the tree and picks the most relevant branches. It doesn’t search everything. It focuses only on the sections that seem most useful first.
VectifyAI released an open-source tool called PageIndex in September 2025. It lets developers submit PDFs through an API. The tool builds the tree automatically and organizes the document into sections with summaries. Developers can find a hands-on guide in the PageIndex GitHub repo, which includes a Vectorless RAG Notebook.
The retrieval process works in two steps. First, the AI looks at the tree without the full text and picks the node IDs that look most relevant. It returns a JSON response with its reasoning and a list of nodes.
Second, the system pulls the full content from those nodes and sends it to the AI to generate a final answer. The AI is told to answer only from that content.
Testing showed strong results on a benchmark called FinanceBench. The PageIndex system scored 98.7% accuracy. Traditional vector RAG scored around 50%. Developers who ran four Python experiments found that most non-PageIndex methods didn’t perform well.
The gap in performance comes down to how each system finds information. Vector RAG uses similarity scores, which don’t always match what’s actually relevant. Vectorless RAG uses summaries and titles to make logical choices about where to look. This matters most in complex documents like financial reports, where similar-sounding words don’t always lead to the right answers. Bluetooth technology faced a similar evolution, as Bluetooth Low Energy was introduced in version 4.0 to address efficiency challenges in devices that needed to communicate smarter, not harder.
The code uses Python libraries including PageIndexClient, asyncio, and requests. Unlike traditional methods, vectorless RAG produces traceable retrieval decisions that make it easier for developers to understand and verify how answers were reached. PageIndex operates without embeddings, chunking, or a vector database, relying instead on LLM-powered reasoning to navigate and extract information from structured documents.
References
- https://www.geeksforgeeks.org/artificial-intelligence/vectorless-rag-pageindex/
- https://www.buildfastwithai.com/blogs/vectorless-rag-pageindex-guide
- https://www.youtube.com/watch?v=97GkSYzr6yk
- https://pyshine.com/PageIndex-Vectorless-Reasoning-RAG/
- https://github.com/VectifyAI/PageIndex
- https://gaodalie.substack.com/p/rag-is-not-dead-no-chunking-no-vectors
- https://pub.towardsai.net/vectorless-rag-how-i-built-a-rag-system-without-embeddings-databases-or-vector-similarity-efccf21e42ff
