treewalk navigation

treewalk is the engine's canonical navigation strategy. It turns retrieval into a navigation problem: an LLM agent starts at the root of the document tree and walks toward the nodes that answer the question.

treewalk is the only name for this strategy. The earlier name "pageindex" is retired — do not use it in code, configs, or docs.

How a walk works

Start at the root

The agent sees the top-level structure of the tree — the titles of the document's main sections — and the question.

Decide where to go

The agent reasons about which branch is most likely to contain the answer and descends into it, optionally expanding a node to read its content.

Walk until confident

It repeats: read titles, expand promising nodes, move up or down. Branches that clearly don't matter are never expanded — keeping the work focused.

Answer and cite

When the agent has gathered enough, it composes an answer and records the exact nodes it relied on. Those become the citations.

Explainable. Each step is a reasoned decision, not a distance score.
Structure-aware. The agent uses the document's own hierarchy as a guide.
Self-correcting. If a branch turns out to be a dead end, the agent can back out and try another — something a single similarity query cannot do.

Using treewalk

You select it per request:

const result = await vl.ask({
  document: doc.id,
  question: 'How is revenue recognized for multi-year contracts?',
  strategy: 'treewalk',
});

The result includes the answer, the citations, and (when enabled) the path the agent took through the tree — useful for debugging and for trust.

treewalk spends LLM calls navigating instead of spending storage on embeddings. Tuning how deep and how broad it walks is covered in the engine configuration reference (coming with the API reference).