Vectorless RAG (reasoning-based retrieval)
Vectorless RAG is retrieval-augmented generation without an embedding index: instead of fetching the chunks whose vectors are closest to the question, the model decides what to read by reasoning over the material’s structure — a table of contents, section headings, a summary of each file — and then reads the parts it chose.
- Also called
- Reasoning-based retrieval, tree-based retrieval
- Related
- Agentic RAG, full-text-first
- Trade
- Better choices, more model calls
What changes when the vectors go
Standard RAG decides what the model sees with a number: the similarity between the question’s embedding and each chunk’s. Whatever ranks highest is read; everything else is not. Nobody — not the model, not you — looks at the material and chooses.
Vectorless RAG moves that decision to the model. It is shown the shape of the material first — chapter headings, a section map, a one-line gist of each document in a folder — and decides which parts bear on the question, opens those, and reads them. If what it read points somewhere else, it goes back for more. That is closer to what a person does with a book they have not read: look at the contents, turn to the chapter, read it properly.
In MindMax, finding candidate passages for a lookup is done by the words they share with the question; reading the structure and choosing what to open is the model’s part.
Same question, two deciders. On the left a score picks the fragments; on the right the model reads the map and picks the sections.
What it gains
- Structure survives. Chunking cuts a book into paragraph-sized pieces with no chapters and no order. Navigating by the table of contents keeps both, which is what questions about an argument depend on.
- Whole sections, not slivers. Having chosen a section, the model reads it in full, so meaning is not cut off at an arbitrary chunk boundary.
- Exact terms stay exact. A clause number or a name is matched as written instead of being blurred into “things that look similar” — see semantic search.
- Reading is on record. Because the model chose what to open, the answer can list what it actually read, rather than what happened to score well.
- No index to build. No embedding pass before the first question, and nothing to re-embed when a file changes.
What it still costs
- Time and model calls. Reading a map and choosing is work a vector lookup does in milliseconds. Answers take longer, and each choice is a model call you pay for.
- It is only as good as the map. A document with no usable headings — a long transcript, a scanned book with no contents page — gives the model little to reason over.
- Something still has to find candidates. Across thousands of files, the model cannot read every summary. Practical systems combine structure with a fast search underneath, and that search has its own blind spots.
How MindMax handles it
MindMax is vectorless by design. There is no vector database behind the search. It finds material the way you would — by structure, by the words on the page, and by trying other wordings — so each answer traces to a passage you can open. The trade-off: a question phrased very differently from your files, or asked in another language, can miss.
What the model reasons over depends on the size of the material. A document that fits the reading budget is not retrieved from at all — it is read whole (full-text-first). A book is navigated by its chapter map, and the sections the question needs are read in full. A question about a folder starts from a one-line gist of every document, picks the ones it needs, and reads them the same way. Looking for an exact sentence is done as a lookup, by the words on the page.
To be precise about the parts: finding candidate passages is word-based, and it is the same search you use in the search panel. Choosing — which documents, which sections, which passages fit what you meant — is the model’s reasoning. And the answer ends with what was actually read for it: a document only counts as read if it was opened or quoted.
- PDFanalyst-report-2025.pdfSizes the market top-down, from shipments
- DOCXstrategy-memo.docxBuilds a bottom-up estimate from customers
- PPTXkickoff-deck.pptxQuotes one headline figure, no method
- XLSXpricing-2024.xlsxPrice list by tier
- EPUBcrossing-the-chasm.epubAdoption stages, in general
A question about a folder: gists first, then the documents the question needs, then their sections in full — and a record of what was read.
A question about one chapter of a whole book: the section was opened and read, and the answer quotes it.
Common questions
What is vectorless RAG?
Retrieval-augmented generation without an embedding index. Instead of retrieving the chunks most similar to the question, the model navigates the material’s structure — headings, a table of contents, per-file summaries — decides what to read, and reads it.
Is vectorless RAG better than vector search?
It is better at keeping structure, exact terms and a record of what was read, and it needs no index. It is slower, uses more model calls, and depends on the material having usable structure. For huge collections, a fast search underneath is still needed.
What is the difference between vectorless RAG and agentic RAG?
Agentic RAG is the broader idea of a model running retrieval itself over several steps, with whatever tools it has — which can include vector search. Vectorless RAG is the narrower choice to do without embeddings and decide by reasoning over structure.
Does MindMax use vectorless RAG?
Yes. There is no vector database. A document that fits is read whole; a book is navigated by its chapters; a folder starts from a gist of each file. Candidate passages are found by their words, and the model decides what to open. The trade-off: a question worded very differently from your files, or in another language, can miss.