pdf_text

Positioned PDF text extraction primitives for semantic PDF comparison.

pdf_text consumes interpreted content stream operations and emits positioned text runs with glyph evidence, normalized text, bounding boxes, provenance, and marked-content references. It is the text-extraction stage used by semantic-pdf-diff before layout clustering and semantic diffing.

What This Crate Provides

• Extraction of text from Tj and TJ operations.
• Basic text matrix, text line movement, leading, character spacing, word spacing, and horizontal scaling support.
• Glyph token evidence, raw text bytes, normalized text, and approximate bounding boxes.
• Marked-content tag and /MCID preservation for tagged-PDF mapping.
• A public font resource model with resource name, object ID, subtype, base-font, encoding, simple-font /FirstChar plus /Widths metrics, descendant-font, and /ToUnicode references.
• Public /ToUnicode CMap parsing and application helpers for bfchar, controlled bfrange, and multi-byte text bytes.
• Conservative Base14 Latin fallback for simple Type1 Helvetica, Times, and Courier-family fonts when shown bytes are printable ASCII/basic whitespace and no custom encoding is present.
• Whitespace normalization for deterministic downstream matching.
• MISSING_TOUNICODE diagnostics when extraction cannot use /ToUnicode or the conservative Base14 Latin fallback safely.

Pipeline Context

pdf_text sits between content interpretation and semantic layout:

pdf_content ContentProgram -> pdf_text TextRun values -> pdf_semantic nodes

The current CLI uses this crate's font resource model, /ToUnicode helpers, and safe Base14 Latin fallback before text-run extraction. Broader font decoding, CID /W metrics, vertical writing, and shaping remain planned work.

Current Compatibility Boundary

This crate is designed for digitally generated PDFs with extractable text. It does not perform OCR, full font shaping, real glyph outline measurement, or renderer-grade text positioning. Image-only PDFs should use the CLI's external OCR adapter path, not this crate as a substitute for OCR.

Bounding boxes are approximate extraction evidence for diff reports, not a pixel-accurate rendering contract.