Player not loading? Watch on YouTube
Krish Naik explains vectorless RAG through PageIndex and a Python notebook. He compares conventional retrieval using chunks, embeddings and vector similarity with a document tree that contains section titles, page references and summaries. In the approach he describes, an LLM selects relevant nodes from the JSON index, then uses their content to answer a question without a vector database.
The explanation covers PDFs with and without a table of contents. Naik says that when a document lacks one, the LLM infers headings and structure, dividing content at section boundaries rather than fixed token counts. He also demonstrates the PageIndex chat interface before moving to the notebook.
The practical setup uses the PageIndex SDK, an OpenAI client and dotenv to load API keys. The notebook uploads a syllabus PDF, polls the indexing status and retrieves the tree with node summaries. It then asks about modern LLM fine-tuning, selects node IDs and passes the matching sections to an answer prompt that requests section titles and page citations.
This example uses hosted APIs; it does not demonstrate an offline or self-hosted deployment. Naik leaves storage and memory handling for large trees to a future discussion. His comparisons explain the retrieval design, but the tutorial does not establish that it outperforms vector search across documents.
