Vectorless
Core Concepts

treewalk, the generative fallback

The strategy in which a chat model walks the contents and pages, used when Jev is unavailable.

Updated 2026

treewalk is the engine's generative navigation strategy. Before navigation moved to Jev in September 2026 it was the main strategy. It is now the fallback.

How a walk works

A chat model gets three tools and up to eight turns:

  1. get_document_structure() returns the contents tree: titles and page ranges, no body text.
  2. get_pages(start_page, end_page) returns the text of those pages.
  3. done(answer, cited_pages, reasoning) ends the walk with the answer and the page ranges it cites.

The model writes its answer inside the done call, so there is no separate answer step.

When it runs

  • No TypeSafe key. judgewalk needs Jev; without it the engine uses treewalk instead.
  • Jev refuses a request. No credit, a rejected key, or an outage: the server binary serves that query with treewalk and logs retrieval: primary model unavailable, served by fallback.
  • On request. POST /v1/answer/treewalk always uses it, and retrieval.strategy: treewalk makes it the default.

How it compares

treewalk makes several generative calls per question where judgewalk makes none before the answer. In the 2026-10-08 comparison on the 3M 2018 10-K, treewalk on glm-4.5-air answered in 7.6 s at the median against 10.5 s for PageIndex on the same model, and PageIndex was faster on 2 of the 6 questions. The shipped Jev path answered in 5.27 s. See Benchmarks.

Pincites are not attached to treewalk answers; it returns cited page ranges.

On this page