0.3.0 — 2026-07-26

Release Notes

Added

• --ocr, an opt-in build feature (#3). deepdoc --ocr scan.pdf recognizes a scan and
  prints Markdown in one command instead of two — for builds that ask for it. It is off by
  default and absent from every released binary: the models alone are 12 MB against a 3.8 MB
  binary, cargo-dist builds one artifact set for all channels (so "brew with models, crates.io
  without" is not expressible), and recognition is probabilistic, which is not something to put
  on the default path of a deterministic parser. Two features, so the cost is the one you choose:

  Feature	What it adds
  ocr	Recognition; models come from --ocr-model <dir> or DEEPOCR_MODEL_DIR. No weights, no network stack.
  ocr-fetch-models	…plus downloading them to the user cache on first run — the only build of DeepDoc that touches the network.

  There is deliberately no ocr-embed-models. deepocr-core can compile its weights in, but it
  reads them with include_bytes! from a directory DeepOCR's own CI populates before building;
  the crate published to crates.io does not carry them, so the feature cannot compile for anyone
  installing DeepDoc normally. A build flag that only works inside one repository's CI is a trap.

  Recognition runs entirely in memory: the page becomes a searchable PDF that never reaches the
  filesystem, then goes through DeepDoc's ordinary PDF reader, so an OCR'd scan gets the same
  reading-order, column and heading handling as any born-digital file. A standalone .png,
  .jpg or .tiff also becomes readable — not a document to DeepDoc, but a page to DeepOCR.

  --ocr is an extra attempt, never a new failure mode: when recognition cannot help, the
  original verdict stands, so a batch skips exactly what it skipped before, for the same reason.
  It only triggers on a file that yields no text at all, so a mostly-text PDF with one scanned
  page inside still extracts what it has and leaves that page alone.

• "ocr": true in the manifest on files whose text was recognized rather than read. Absent on
  every other row, so a run without --ocr emits exactly the 0.2.0 schema. Deterministically
  parsed text and OCR output are not the same evidence — and a chunk hash that changed because
  a page was recognized differently is a different story from one that changed because the parser
  improved.

Changed

• --ocr now explains itself in every build. It used to fail with clap's
  unexpected argument '--ocr', which is true and useless. The flag is defined everywhere; a
  build without the feature rejects it with the feature to install and the two-tool loop that
  needs no special build. Same exit code (2).
• The dependency-graph promise is now explicitly about the default build. With the ocr
  feature the graph gains deepocr-core and its tree, including dirs-sys (pure Rust despite the
  name — libc and option-ext, nothing compiled) and hayro, a page rasterizer. Rendering
  pages is otherwise deliberately out of scope here; behind an off-by-default flag it is an
  intentional exception. The default build is unchanged and still contains no C or -sys crates.

Notes

• Determinism is suspended for pages that go through recognition, and a run using --ocr says so
  on stderr. Everything else remains byte-for-byte reproducible.
• The ocr feature needs a newer toolchain than the default build (deepocr-core requires Rust
  1.92; DeepDoc itself still builds on 1.85).

Install deepdoc 0.3.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/deeplabua/deepdoc/releases/download/v0.3.0/deepdoc-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/deeplabua/deepdoc/releases/download/v0.3.0/deepdoc-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install deeplabua/tap/deepdoc

Download deepdoc 0.3.0

File	Platform	Checksum
deepdoc-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
deepdoc-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
deepdoc-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
deepdoc-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

0.2.0 — 2026-07-26

Release Notes

Auditable ingestion: a parser upgrade now says which chunks it actually changed, and a batch now
answers "what happened to every file?" in JSON instead of prose on stderr. Both came from
engineers reading the launch write-up (#1, #2).

Added

• A content hash on every chunk (#1). --chunk --format json gives each chunk a
  hash: "sha256:<hex>", so re-extracting a corpus after a parser upgrade tells you which chunks
  moved and which can keep their embeddings. The hash covers the chunk's heading path and text
  together (heading_path.join("\u{1F}") + "\u{1E}" + text, the ASCII unit and record
  separators as the boundary): a chunk only contains its own heading when it starts with one, so
  for everything else the context lives outside text, and re-filing the same sentences under a
  new heading — the exact case worth catching — would otherwise hash identically.

• --manifest <path> (#2): one JSON array for the whole run naming, per input, its
  status (extracted / skipped / error), the reason there is no output (no_text_layer,
  unsupported_format, parse_error, io_error), the detected format, and the output file
  that was actually written. Exit code 4 is a process-level signal, so routing scans to OCR
  meant either grepping stderr or hand-rolling a loop over the files — and a hand-rolled loop
  re-derives the output names, where report.pdf and report.docx collapse into one
  report.md. output reports the name the batch really chose (out/report.pdf.md), so the
  pipeline never computes it. Inputs that could not even be listed get an entry too. It works
  without --recursive (an array of one), is deterministically ordered, and is a report, not a
  policy: it changes no exit code and turns no skip into a failure.

  deepdoc ./corpus --recursive -o out/ --manifest run.json
  jq -r '.[] | select(.reason=="no_text_layer") | .source' run.json | xargs -n1 deepocr

Changed

• Breaking for deepdoc-core users: Chunk has a new public field, hash, so code that
  builds one with a struct literal (Chunk { text, heading_path, source, byte_range }) no longer
  compiles. Add hash: deepdoc_core::chunk::chunk_hash(&heading_path, &text). Reading chunks —
  the common case — is unaffected.
• The chunk header in --format md / --format text now carries a short form of the hash, so
  "which chunk is this?" is answerable in every format, not only in JSON:
  <!-- chunk 2/5 | Handbook > Payroll | bytes 296-390 | sha256:2b7c1d0f8a3e… -->.

Notes

• sha2 (RustCrypto) does the hashing: pure Rust, MIT OR Apache-2.0, and already in the graph
  behind the PDF reader — the dependency tree gained no new crates, and still contains no C or
  -sys crates.

Install deepdoc 0.2.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/deeplabua/deepdoc/releases/download/v0.2.0/deepdoc-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/deeplabua/deepdoc/releases/download/v0.2.0/deepdoc-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install deeplabua/tap/deepdoc

Download deepdoc 0.2.0

File	Platform	Checksum
deepdoc-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
deepdoc-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
deepdoc-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
deepdoc-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

0.1.1 — 2026-07-26

Release Notes

Changed

• A document with no extractable text now points at the tool that fixes it. The message used to
  say a scan "needs --ocr" — a flag DeepDoc does not have, and rejects with exit 2. It now names
  DeepOCR, which writes an invisible text layer over the
  page image; the result is born-digital and extracts normally. Exit code 4 is unchanged, so an
  ingestion pipeline keeps its routing signal.
• The README documents that loop under Scanned documents. OCR stays deliberately outside this
  parser: recognition is probabilistic, and mixing it in would cost the determinism DeepDoc is
  built on. A --ocr feature that links DeepOCR's engine in directly remains planned and off by
  default, so the released binary stays small and model-free.

Install deepdoc 0.1.1

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/deeplabua/deepdoc/releases/download/v0.1.1/deepdoc-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/deeplabua/deepdoc/releases/download/v0.1.1/deepdoc-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install deeplabua/tap/deepdoc

Download deepdoc 0.1.1

File	Platform	Checksum
deepdoc-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
deepdoc-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
deepdoc-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
deepdoc-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

v0.1.0

Install deepdoc 0.1.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/deeplabua/deepdoc/releases/download/v0.1.0/deepdoc-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/deeplabua/deepdoc/releases/download/v0.1.0/deepdoc-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install deeplabua/tap/deepdoc

Download deepdoc 0.1.0

File	Platform	Checksum
deepdoc-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
deepdoc-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
deepdoc-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
deepdoc-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum