Struct BboxBook 

pub struct BboxBook {
    pub tags: Vec<i64>,
    pub otsl: Vec<i64>,
    pub hiddens: Vec<f32>,
    pub n: usize,
    pub merge: HashMap<usize, i64>,
    /* private fields */
}

The autoregressive loop’s per-step state, mirroring docling’s predict bookkeeping (tag_H_buf / bboxes_to_merge): which decoder hidden states feed the bbox decoder and how horizontal spans merge. Both the native and browser loops step the decoder themselves (sync vs async ort) and feed each result through step; everything else stays here so the two paths can’t drift.

Fields

tags: Vec<i64>

The decoder input prefix ([START], then every emitted tag).

otsl: Vec<i64>

The emitted OTSL structure tokens (no START/END).

hiddens: Vec<f32>

Per-bbox-cell decoder hidden states, flattened [n, EMBED_DIM].

n: usize

Number of hidden states collected (hiddens.len() / EMBED_DIM).

merge: HashMap<usize, i64>

Span merges: cur_bbox_ind → partner (-1 → the last box).

Implementations

impl BboxBook

pub fn new() -> Self

pub fn step(&mut self, raw: i64, hidden: &[f32]) -> bool

Feed one raw decoded tag and its hidden state. Returns false to stop decoding: when the corrected tag is END (not recorded), or when the current row reached MAX_ROW_TAGS (runaway).

pub fn runaway(&self) -> bool

The decode degenerated into one endless row (no nl for MAX_ROW_TAGS tags). docling keeps such a sequence, but its span recovery (html_to_otsl knows colspans 2–20 only) and orphan pickup then drop most of the table’s words; callers reject the structure so the region takes the geometric fallback instead. A long table that merely hits MAX_STEPS with ordinary rows is not a runaway.

Trait Implementations

impl Default for BboxBook

fn default() -> Self

Returns the “default value” for a type.