glyphive

Archive a file tree to compact, OCR-friendly, printable pages and restore it from a transcript or scan. Glyphive keeps the paper representation human-readable while using checksums and error correction to detect or repair the character errors that make ordinary base64-on-paper fragile.

Status: alpha. The wire format and CLI may change before 1.0.

Features

• Printable text, PDF, or Word output with selectable font family and size.
• Bundled OCR-B PDF option under the SIL Open Font License, plus custom .ttf/.otf PDF font paths for measured channels.
• Measured OCR-safe base16g-crc16-rs alphabet (ABCDHKLMPRTVXY34) with no confusable character aliases — and no trained OCR model required, for any codec.
• Localized integrity checks on every encoded line and protected page metadata.
• Document-wide Reed-Solomon parity for correcting scattered OCR errors.
• Optional whole-page recovery (--parity-pages K, default off): survives up to K wholly lost/unscannable pages, independent of the per-line parity.
• Binary-safe archives for nested trees, empty directories, and arbitrary file contents.
• Deterministic restore from text; optional OCR providers are loaded only when image input is requested.

Installation

pip install glyphive

Optional features:

Glyphive requires pathlib_next>=0.9.0,<0.10, duho>=0.6.0,<0.7, and Python 3.9 or newer.

Quick start

Create a text archive from one directory, inspect it, and restore it:

glyphive create -f backup.txt --compression gzip -C project .
glyphive list -f backup.txt
glyphive inspect -f backup.txt          # recovery-headroom report (read-only)
glyphive extract -f backup.txt -C restored
glyphive info                           # what codecs/OCR engines/fonts are available

create picks its page geometry from a measured preset, --mode {conservative,standard,max}. standard is the default: the base16g-crc16-rs codec, the bundled dejavu-sans-mono font, 6pt, --line-width max, and regular margins — the most blur-tolerant combination measured on a real create → rasterize → OCR → extract → diff restore gate. conservative backs off to 8pt with --line-width auto (the OCR-measured-safe cap, ≤ 60 characters); max keeps standard's settings but uses minimal margins for the smallest page count. Any of --codec/--font/--font-size/--line-width/--minimal-margins passed explicitly overrides just that one field of the preset. Documents created before 0.3.0 used 11pt and --line-width auto, so their geometry differs from what a bare create produces now. Full table and the measurements behind it: the create guide's preset section.

Restore or inspect an already-generated GQ1 QR image set explicitly with glyphive extract -f qr-pages/ --from-qr -C restored or glyphive list -f qr-pages/ --from-qr. Ordinary image input continues through OCR; QR mode requires glyphive[qr] and rejects mixed, duplicate, corrupt, or incomplete symbol sets before writing files.

Create QR-only or hybrid human-readable/QR PDF pages with an explicit format:

glyphive create -f backup.pdf --format qr -C project .
glyphive create -f backup-hybrid.pdf --format hybrid -C project .

A .pdf filename without --format remains the ordinary text PDF renderer; the suffix cannot distinguish PDF, QR, and hybrid presentations.

Tar-style mode flags are equivalent when a positional command is inconvenient:

glyphive -c -f backup.txt -C project .
glyphive -t -f backup.txt
glyphive -x -f backup.txt -C restored

Installed plugins

Glyphive can explicitly discover implementations supplied by installed Python distributions. Pass the global --plugins flag to opt in for that invocation:

glyphive --plugins create -f backup.txt --codec vendor_codec -C project .

Plugin distributions register a concrete typed class under one of these entry point groups:

The entry-point name must exactly match the class's lowercase registry name. Library callers use glyphive.plugins.discover(); normal imports and registry lookups never discover plugins. Discovery is deterministic and cached, and a bad candidate is reported without preventing valid candidates from loading.

Installed plugin code executes with the same permissions as Glyphive and is not sandboxed. Use --plugins only when you trust every installed distribution that declares one of these groups. Discovery does not download or update code.

Create a PDF instead:

pip install "glyphive[pdf,zstd]"
glyphive create -f backup.pdf --compression zstd -C project .

Creation uses bounded-memory disk spools; --temp-dir selects their location and --chunk-size tunes sequential I/O for unusually constrained systems.

Restore scans or generated documents with Tesseract. Input type is detected from file contents first and the extension second:

pip install "glyphive[ocr,document-input]"
glyphive extract -f scans/ --ocr-engine tesseract -C restored
glyphive list -f backup.pdf --ocr-engine tesseract

Restore verifies global integrity before publishing staged files. Advanced resource controls include --temp-dir, --chunk-size, and --max-output-bytes on both extract and list. Transcript frames, compressed payload, decompressed archive, and staged files are disk-backed; Reed-Solomon correction retains only compact offsets and one codeword at a time.

The operating-system Tesseract executable and language data must also be installed. PDF input uses pypdfium2; Glyphive-generated DOCX transcripts are read directly with python-docx, without Microsoft Word or LibreOffice. See the create guide, restore guide, and OCR guide for details.

Format at a glance

The default base16g-crc16-rs format uses exactly 16 payload symbols, or 4 bits per printed character. Each data (L) or parity (P) line has a masked index and a full CRC-16 encoded in the same safe alphabet:

L<5 safe index chars> <up to 60 safe payload chars> #<4 safe CRC chars>

CRC-protected H header frames carry the codec, compression method, page count, and whole-document SHA-256. Each page has a protected T footer with its page number and a truncated page hash. Human-facing #!glyphive and PAGE n/total text is display-only; restore trusts the protected frames.

Which codec should I use?

base16g-crc16-rs (the default), with any font. It restores byte-for-byte on stock Tesseract from 8pt down to 4pt — and 3pt with OCR-B — so the cheapest way to get more data per page is a smaller font, not a wider alphabet: 4pt is roughly 4× denser than 8pt.

base32g-crc16-rs is ~25% denser per character and restores on stock too, but only with Courier (measured 4–10pt). It fails on OCR-B and DejaVu Sans Mono at every size, because the punctuation it adds (?@!&+=) gets dropped or mangled on those fonts, and a dropped glyph shortens the line in a way Reed-Solomon cannot repair. The 64-glyph codecs are encode-only: no conflict-free 64-glyph set exists in printable ASCII.

No trained OCR model is needed — or recommended — for any of them.

Small, scattered line errors become known erasures and can be repaired by the document-wide Reed-Solomon parity. A missing page is reported rather than guessed, and the document-wide RS may still recover it from the surviving pages when the budget allows. create --parity-pages K (default 0, off) opts into a separate, guaranteed layer: K extra pages of document-level parity that survive up to K wholly lost pages regardless of the per-line budget. Read the wire-format guide for the complete layering and framing rules.

Standalone zipapp

Releases may include a self-contained glyphive.pyz. Build the universal core artifact with:

python package.py --out dist/glyphive.pyz
python dist/glyphive.pyz --help

The universal zipapp includes text output with none and gzip compression. Optional, platform-specific artifacts can be built with --extras; use python package.py --help for the declared choices.

API overview

Full generated API documentation is available on the documentation site.

Development

See CONTRIBUTING.md for setup, test, documentation, and pull request guidance.

Use a project-local virtual environment, then run the lightweight suite:

python -m venv .venv/dev
.venv/dev/Scripts/python -m pip install -e ".[all,dev,docs]"
.venv/dev/Scripts/python -m pytest -q
.venv/dev/Scripts/python -m mkdocs build --strict

On POSIX, replace .venv/dev/Scripts/python with .venv/dev/bin/python. OCR sweeps and performance measurements are separate manual workloads; see Benchmarks.

Releasing

Glyphive follows Semantic Versioning and keeps a CHANGELOG.md. Pushing a v* tag runs the release validation, package build, publication, and documentation deployment workflow.

License

MIT — see LICENSE.

The bundled OCR-B font retains its SIL Open Font License 1.1; see third-party licenses.