Tree retrieval

A document is not a bag of words — it has structure. Titles, sections, sub-sections, lists, and tables encode the author's own map of the content. Vectorless treats that structure as the retrieval index.

From document to tree

When you ingest a document, the engine parses it into a tree of nodes. Each node corresponds to a structural unit of the source:

Annual Report
├─ 1  Overview
├─ 2  Strategy
│  ├─ 2.1  Market
│  └─ 2.2  Roadmap
└─ 3  Financials
   ├─ 3.1  Q2 Results
   └─ 3.2  Q3 Results
      └─ 3.2.1  Revenue

Every node knows:

• its title and content,
• its path from the root (e.g. Financials → Q3 Results → Revenue),
• its parent and children.

Why a tree

• Locality. Related content stays together, the way the author intended.
• Addressability. Any answer can be named by its path — that is what makes citations exact.
• Navigability. An agent can move up, down, and across the tree, expanding only the branches that matter — see treewalk.

What this replaces

This single representation does the job that chunking, embedding, and a vector index do in classic RAG — without any of those moving parts. The no-chunking model explains why.