pdfmuse_core/

ir.rs

//! Unified intermediate representation (IR).
//!
//! Every source format (PDF, DOCX) is parsed down to this one representation, and
//! every binding (Python/Node/WASM) serializes it through the same serde path —
//! that shared path is the technical guarantee behind byte-identical output.
//!
//! **Coordinate convention:** all coordinates are normalized to a **top-left
//! origin, Y growing downward, unit = pt**. PDF's native bottom-left origin is
//! converted at the parsing edge and never leaks into the IR, so every downstream
//! consumer sees one consistent coordinate space.
//!
//! `Char` is kept at the finest granularity on purpose: it lets downstream users
//! bypass our clustering and regroup by coordinate themselves — we never subtract
//! for the caller. Fields here mirror the technical plan (§4).

use serde::Serialize;

/// A fully parsed document — the root of the IR.
#[derive(Serialize, Clone, Debug, Default)]
pub struct Document {
    /// Which source format produced this document.
    pub source: SourceKind,
    pub metadata: Metadata,
    pub pages: Vec<Page>,
    /// Bookmarks / table of contents.
    pub outline: Vec<OutlineItem>,
    /// Non-fatal degradations recorded during parsing; parsing is never aborted
    /// for these (see the "graceful degradation" principle).
    pub warnings: Vec<Warning>,
}

/// The format a [`Document`] was parsed from.
#[derive(Serialize, Clone, Debug, Default, PartialEq, Eq)]
pub enum SourceKind {
    #[default]
    Pdf,
    Docx,
}

/// Document-level metadata. All fields optional — absent in the source ⇒ `None`.
#[derive(Serialize, Clone, Debug, Default)]
pub struct Metadata {
    pub title: Option<String>,
    pub author: Option<String>,
    pub subject: Option<String>,
    pub keywords: Option<String>,
    pub creator: Option<String>,
    pub producer: Option<String>,
    pub page_count: u32,
}

/// A single page and everything found on it, from finest (chars) to coarsest
/// (blocks). Layers coexist so callers can pick the granularity they need.
#[derive(Serialize, Clone, Debug, Default)]
pub struct Page {
    pub index: u32,
    pub width: f32,
    pub height: f32,
    /// Page rotation in degrees (0/90/180/270).
    pub rotation: i32,
    /// Finest granularity: every glyph with its coordinates.
    pub chars: Vec<Char>,
    /// Characters clustered into lines (geometric, deterministic).
    pub lines: Vec<TextLine>,
    /// Paragraphs / tables / images in reading order.
    pub blocks: Vec<Block>,
    /// Vector rectangles — a source of table borders.
    pub rects: Vec<Rect>,
    /// Vector line segments — a source of table rules.
    pub rules: Vec<Rule>,
    pub images: Vec<ImageRef>,
    pub links: Vec<Link>,
}

/// A single character with precise placement. `text` is always Unicode
/// (post-CMap) — never a raw CID.
#[derive(Serialize, Clone, Debug)]
pub struct Char {
    pub text: String,
    pub bbox: BBox,
    pub font: FontRef,
    pub size: f32,
    /// RGB in 0.0..=1.0, if known.
    pub color: Option<[f32; 3]>,
}

/// An axis-aligned bounding box in normalized (top-left origin, Y-down, pt) space.
#[derive(Serialize, Clone, Copy, Debug, Default, PartialEq)]
pub struct BBox {
    pub x0: f32,
    pub y0: f32,
    pub x1: f32,
    pub y1: f32,
}

/// Reference to a font by resource name (details filled in by `fonts.rs`, PER-37).
#[derive(Serialize, Clone, Debug, Default)]
pub struct FontRef {
    pub name: String,
}

/// A line of text produced by clustering [`Char`]s.
#[derive(Serialize, Clone, Debug)]
pub struct TextLine {
    pub bbox: BBox,
    pub text: String,
    /// Indices into the owning [`Page::chars`] that make up this line.
    pub chars: Vec<u32>,
}

/// A coarse-grained page element in reading order.
#[derive(Serialize, Clone, Debug)]
pub enum Block {
    Paragraph(Paragraph),
    Table(Table),
    Image(ImageRef),
}

/// A paragraph of text. `heading_level` is `Some(n)` when this paragraph is a
/// heading (drives Markdown `#` levels and RAG `heading_path`).
#[derive(Serialize, Clone, Debug)]
pub struct Paragraph {
    pub bbox: BBox,
    pub text: String,
    pub heading_level: Option<u8>,
}

/// A reconstructed table. `source` records which deterministic path built it.
#[derive(Serialize, Clone, Debug)]
pub struct Table {
    pub bbox: BBox,
    /// Row-major grid of cells.
    pub rows: Vec<Vec<Cell>>,
    pub source: TableSource,
}

/// Which deterministic reconstruction path produced a [`Table`].
#[derive(Serialize, Clone, Debug, PartialEq, Eq)]
pub enum TableSource {
    /// Built from vector rules/rects (highest precision).
    Ruled,
    /// Built from whitespace-aligned text columns (only above a confidence threshold).
    Whitespace,
    /// Explicit table structure from a DOCX `w:tbl`.
    Docx,
}

/// A single table cell, possibly spanning rows/columns.
#[derive(Serialize, Clone, Debug)]
pub struct Cell {
    pub text: String,
    pub bbox: BBox,
    pub row_span: u16,
    pub col_span: u16,
}

/// A vector rectangle (e.g. a table border box).
#[derive(Serialize, Clone, Debug)]
pub struct Rect {
    pub bbox: BBox,
}

/// A vector line segment (e.g. a table rule).
#[derive(Serialize, Clone, Debug)]
pub struct Rule {
    pub x0: f32,
    pub y0: f32,
    pub x1: f32,
    pub y1: f32,
    /// Stroke width in pt.
    pub width: f32,
}

/// A reference to an embedded image and where it sits on the page.
#[derive(Serialize, Clone, Debug)]
pub struct ImageRef {
    pub id: String,
    pub bbox: BBox,
    pub width: u32,
    pub height: u32,
}

/// A hyperlink region. Exactly one of `uri` / `page` is typically set
/// (external link vs. intra-document jump).
#[derive(Serialize, Clone, Debug)]
pub struct Link {
    pub bbox: BBox,
    pub uri: Option<String>,
    pub page: Option<u32>,
}

/// An entry in the document outline (bookmarks / TOC). Recursive.
#[derive(Serialize, Clone, Debug)]
pub struct OutlineItem {
    pub title: String,
    pub page: Option<u32>,
    pub level: u8,
    pub children: Vec<OutlineItem>,
}

/// A non-fatal degradation recorded during parsing.
#[derive(Serialize, Clone, Debug)]
pub struct Warning {
    /// The page it occurred on, if page-scoped.
    pub page: Option<u32>,
    pub kind: WarningKind,
    pub detail: String,
}

/// The category of a [`Warning`].
#[derive(Serialize, Clone, Debug, PartialEq, Eq)]
pub enum WarningKind {
    /// An object could not be parsed and was skipped (see PER-38).
    MalformedObject,
    /// A CID font lacked a usable CMap/ToUnicode mapping.
    MissingCMap,
    /// The document was decrypted via a fallback path.
    EncryptedFallback,
    /// A scanned page has no text layer and needs OCR (pluggable backend).
    NeedsOcr,
    /// A source feature is not (yet) supported by the core.
    Unsupported,
}

#[cfg(test)]
mod tests {
    use super::*;

    /// The IR must serialize to stable JSON — the foundation of cross-binding parity.
    #[test]
    fn document_serializes_to_stable_json() {
        let doc = Document {
            source: SourceKind::Pdf,
            metadata: Metadata { page_count: 1, ..Default::default() },
            pages: vec![Page {
                index: 0,
                width: 612.0,
                height: 792.0,
                chars: vec![Char {
                    text: "A".to_string(),
                    bbox: BBox { x0: 0.0, y0: 0.0, x1: 10.0, y1: 12.0 },
                    font: FontRef { name: "Helvetica".to_string() },
                    size: 12.0,
                    color: None,
                }],
                ..Default::default()
            }],
            warnings: vec![Warning {
                page: Some(0),
                kind: WarningKind::MissingCMap,
                detail: "font F1 has no ToUnicode".to_string(),
            }],
            ..Default::default()
        };

        // Serializes without error and is deterministic (same input → same string).
        let json = serde_json::to_string(&doc).expect("IR serializes");
        assert_eq!(json, serde_json::to_string(&doc.clone()).unwrap());
        assert!(json.contains("\"source\":\"Pdf\""));
        assert!(json.contains("\"kind\":\"MissingCMap\""));
        assert!(json.contains("\"text\":\"A\""));
    }
}