Vectorless
Core Concepts

Citations

Why every Vectorless answer traces back to exact tree nodes.

Updated 2026

In Vectorless, citations are not a post-hoc "find the source" step bolted onto a generated answer. They are a property of the engine: because the treewalk agent ends its walk on real tree nodes, the nodes it used are the citation.

What a citation contains

Each citation names a node by its place in the tree:

{
  "node": "3.2.1",
  "title": "Revenue",
  "path": ["Financials", "Q3 Results", "Revenue"],
  "snippet": "Q3 revenue rose 12% quarter over quarter ..."
}
  • node — the stable identifier of the node within the document tree.
  • path — the human-readable trail from the root, so a reader knows exactly where in the document the claim came from.
  • snippet — the supporting text the agent relied on.

Why they're trustworthy

Path-correctness

A citation is only useful if it points at the right place. Because citations come from the agent's actual navigation path, they are correct by construction — not reconstructed by a second model guessing at sources.

  • No hallucinated sources. The agent can only cite nodes it actually visited.
  • Granular. Citations land on the specific section or table, not a fuzzy chunk that happened to be nearby in vector space.
  • Auditable. Combined with the walk path, you can see how the engine reached each cited node.

Using citations

Render them as footnotes, link them back into your own viewer using the path, or gate an answer on a minimum number of citations. The SDKs expose them on every response object.

On this page