Introduction
Document retrieval for the reasoning era — no chunking, no embeddings, no vector DB.
Vectorless is a reasoning-based document-retrieval engine. Instead of splitting a document into fixed-size chunks and searching them by vector similarity, Vectorless parses the document into a hierarchical tree and lets an LLM agent navigate it — returning answers with precise, path-correct citations.
The one-line version
Vectorless parses a document into a tree, an agent walks it with the
treewalk strategy, and you get an answer plus the exact nodes it came from.
Why no chunking, embeddings, or vector DB?
Classic RAG turns retrieval into a nearest-neighbor problem: chunk the text, embed every chunk, and hope the question's embedding lands near the right one. That throws away the document's structure and makes citations approximate.
Vectorless takes the opposite bet:
- Structure is signal. A document already has sections, sub-sections, tables, and ordering. Vectorless preserves that as a tree.
- Navigation beats similarity. An LLM agent reasons about where an answer
lives and walks the tree to it — the
treewalkstrategy. - Citations are structural. Because the agent ends on real nodes, every answer maps back to exact locations in the source.
What's in these docs
Getting Started
Install an SDK, point it at a document, and ask your first question.
Core Concepts
Tree retrieval, the no-chunking model, citations, and treewalk.
API Reference
The HTTP API surface — generated from the engine's OpenAPI spec.
SDKs
Official clients for TypeScript, Python, and Go.
Self-hosting
Run the engine yourself on Neon, Cloudflare R2, and Upstash QStash.