pdf-engine 1.2.0

[package]
name = "pdf-engine"
version = "1.2.0"
edition = "2021"
# AGPLv3-or-later, or the PDFluent Commercial Licence. See LICENSE.
license = "AGPL-3.0-only OR LicenseRef-PDFluent-Commercial"
readme = "README.md"
description = "Unified PDF rendering engine — page rendering, text extraction, thumbnails."
# Tests and examples do not travel. They read fixtures from the repository,
# two directories up, which are not in the tarball -- so they cannot build
# outside this workspace, and what they do carry into a published crate is a
# path into a tree the reader does not have. `pdfluent` and `pdf-xfa` have
# kept them out for that reason since they were first published; this one had
# not (#358).
exclude = ["tests/", "examples/"]
repository = "https://github.com/pdfluent/pdfluent-sdk"
homepage = "https://pdfluent.com"
documentation = "https://pdfluent.com/docs"

keywords = ["pdf", "rendering", "text-extraction", "forms"]
categories = ["rendering", "text-processing"]
[lib]
crate-type = ["rlib", "cdylib"]

[dependencies]
pdf-syntax = { path = "../pdf-syntax", version = "=0.6.1" }
pdf-interpret = { path = "../pdf-interpret", version = "=0.6.1" }
pdf-compliance = { path = "../pdf-compliance", version = "=1.2.0" }
lopdf = { package = "pdfluent-lopdf", path = "../lopdf", default-features = false, version = "=0.39.7" }
pdf-render = { path = "../pdf-render", version = "=1.2.0" }
pdf-forms = { package = "pdfluent-forms", path = "../pdf-forms", version = "=1.2.0" }
pdf-xfa = { path = "../pdf-xfa", optional = true, version = "=1.2.0" }
xfa-layout-engine = { path = "../xfa-layout-engine", optional = true, version = "=1.2.0" }
formcalc-interpreter = { path = "../formcalc-interpreter", optional = true, version = "=1.1.0" }
xfa-json = { path = "../xfa-json", optional = true, version = "=1.2.0" }
thiserror = { workspace = true }
kurbo = "0.13.0"
rayon = { version = "1", optional = true }
ocrs = { version = "0.12", optional = true }
rten = { version = "0.24", optional = true }
dirs-sys = { version = "0.3", optional = true }
# Unconditional since the balanced tier (#641): the model specs, the failure
# types and the path resolution are plain Rust in every build, so a build
# without the tier still reports a missing model with the typed `ModelError`
# rather than a string. No `features` here on purpose: `ocr-onnx` wants the
# paddle backend (ort) while `ocr-cli` wants only the system-binary backend
# (std). The features that enable a backend name the backend they need, so
# enabling one never drags in the other's runtime, and the default tree stays
# pure Rust (held by the_default_tree_carries_no_ort.py).
pdf-ocr = { workspace = true }
reqwest = { workspace = true, optional = true }
base64 = { workspace = true }
# The cloud OCR clients encode page images as PNG (ocr.rs); without this the
# library does not build with any ocr-* cloud feature, and only `cargo test`
# (where image is a dev-dependency) hides it.
image = { workspace = true, optional = true }
serde = { workspace = true, optional = true }
serde_json = { workspace = true, optional = true }
hmac = { version = "0.12", optional = true }
sha2 = { version = "0.10", optional = true }
rsa = { version = "0.9", default-features = false, features = ["sha2"], optional = true }
pkcs8 = { version = "0.10", features = ["pem"], optional = true }
time = { version = "0.3", features = ["formatting", "macros"], optional = true }
# Optional ONNX runtime for the vision layer (#636). Licence: MIT OR Apache-2.0
# (see THIRD_PARTY_LICENSES.txt). Off by default; no behaviour change unless
# the `vision` feature is enabled. No default features: the `transformers`
# op-set is not needed for the tiny detection/classification models here and
# would grow the binary for no reason.
tract-onnx = { version = "0.23", optional = true, default-features = false }
# The model tiers' view of the shared ONNX Runtime layer (#641): the same
# `ort` pdf-ocr loads (one per build graph), plus the tensors the model
# features feed it. Optional, behind `ort`; the default tree never resolves
# them (held by `the_default_tree_carries_no_ort.py`). The layout tier builds
# its input tensor from `(shape, Vec<f32>)` and needs no `ndarray` of its
# own; the tables tier does, hence the feature and the dep.
ort = { version = "=2.0.0-rc.13", optional = true, default-features = false, features = ["std", "ndarray", "load-dynamic", "api-23"] }
ndarray = { version = "0.17", optional = true }

[features]
default = []
parallel = ["dep:rayon"]
# Enable serde derives on the canonical wire DTO (`TextSpanInfo`). Off by
# default so the core engine stays serialization-agnostic; bindings that emit
# JSON (Tauri, WASM) turn it on.
serde = ["dep:serde"]
xfa = ["dep:pdf-xfa", "dep:xfa-layout-engine", "dep:formcalc-interpreter", "dep:xfa-json"]
# Opt-in: compile the sandboxed QuickJS runtime into the XFA flatten path so
# `XFA_JS_EXECUTION_MODE=sandboxed` can execute instanceManager scripts. Off by
# default; no behaviour change unless both this feature and the env var are set.
xfa-js-sandboxed = ["xfa", "pdf-xfa/xfa-js-sandboxed"]
ocr = ["dep:ocrs", "dep:rten", "dep:dirs-sys"]
ocr-mistral = ["dep:reqwest", "dep:image", "dep:serde", "dep:serde_json"]
ocr-google = ["dep:reqwest", "dep:image", "dep:serde", "dep:serde_json", "dep:rsa", "dep:pkcs8", "dep:sha2", "dep:time"]
ocr-aws = ["dep:reqwest", "dep:image", "dep:serde", "dep:serde_json", "dep:hmac", "dep:sha2", "dep:time"]
ocr-azure = ["dep:reqwest", "dep:image", "dep:serde", "dep:serde_json"]
ocr-cloud = ["ocr-mistral", "ocr-google", "ocr-aws", "ocr-azure"]
ocr-onnx = ["pdf-ocr/paddle"]
# The shared ONNX Runtime layer, and the two model tiers on top of it (#641).
# Pin (see pdf-ocr): ort =2.0.0-rc.13, api-23, ONNX Runtime 1.23.2, loaded
# dynamically, never linked, never downloaded. `ort-layout` and `ort-tables`
# are empty placeholders here -- their model code lands in later jobs; the
# names are fixed now. Off by default; the default build stays pure Rust.
ort = ["pdf-ocr/onnx-runtime", "dep:ort", "dep:ndarray"]
ort-layout = ["ort"]
ort-tables = ["ort"]
# OCR through the system `tesseract` binary (`TesseractCliEngine`): std only,
# no paddle, no ort. The Python and Node bindings build with this on; where
ocr-cli = ["pdf-ocr/tesseract-cli"]
# Optional, fast vision layer: raster + ONNX runner on tract-onnx (#636).
# Off by default; no trained model ships and default behaviour is unchanged.
vision = ["dep:tract-onnx"]
full = ["parallel", "xfa", "ocr", "ocr-onnx", "ocr-cloud"]
# Research spikes — not compiled in normal CI.  Build explicitly with `--features spike`.
spike = []
# Measurement grids over the corridor guards (#574) — not compiled in normal
# CI. Build explicitly with `--features corridor-grid`. With it on, the
# `PDFLUENT_CORRIDOR` / `PDFLUENT_CORRIDOR_TRACE` overrides are read, so one
# binary serves the whole grid; without it the decoder serves the shipped
# config from the table switch and reads no environment.
corridor-grid = []

[[example]]
name = "chunk_dump"
required-features = ["spike"]

[[example]]
name = "logical_compare"
required-features = ["spike"]

[[example]]
name = "structtree_measure"
required-features = ["spike"]

[[example]]
name = "structtree_spike"
required-features = ["spike"]

[[example]]
name = "structtree_table"
required-features = ["spike"]

[[example]]
name = "structured_demo"
required-features = ["spike"]

[[example]]
name = "table_audit"
required-features = ["spike"]

[[example]]
name = "table_dump"
required-features = ["spike"]

[dev-dependencies]
test-skip = { path = "../test-skip" }
# The render_page_png example writes its PNG with the `png` crate directly, in
# process: the best tier measures these bytes, so they must not depend on a
# converter outside this tree. Already in the lock via `image`.
png = "0.18"
# The Office extractors, for the office_markdown example and its test only.
# Dev-dependencies, because every one of these crates depends on this engine:
# a normal dependency would be a cycle, a dev-dependency is not. The same
# shape as test-skip: versionless path crates the published build never sees.
# The versions must stay off: cargo strips a path-only dev-dependency when it
# packages, but a versioned one has to resolve on the index -- where
# docx-extract and markup-extract never land (publish = false) and the other
# three land only after this engine, so "cargo package" of this crate (and of
# the workspace) fails before and after the publish either way. Measured with
# "cargo package -p pdf-engine" on 03-10-2026: "no matching package named
# docx-extract" with the versions on, a clean package with them off.
docx-extract = { path = "../docx-extract" }
pptx-extract = { path = "../pptx-extract" }
xlsx-extract = { path = "../xlsx-extract" }
markup-extract = { path = "../markup-extract" }
epub-extract = { path = "../epub-extract" }
# In-memory fixture zips for the office_markdown test (same shape as the
# extractors' own dev-dependency).
zip = { version = "9", default-features = false, features = ["deflate"] }
image = { workspace = true }
lopdf = { workspace = true }
# `float_roundtrip`: the learned-model parity tests read LightGBM's feature
# rows from JSON, and the default parser can land one ulp off a value -- which
# is enough to send it to the other side of a split whose threshold IS that
# value, and to fail parity for a model that is right (#599).
serde_json = { workspace = true, features = ["float_roundtrip"] }
# The TATR end-to-end test decodes its fixtures' base64 crop PNGs.
base64 = { workspace = true }
# Dev-only pdfium reference comparator for the render-profiling milestone.
# Binds at runtime to a local libpdfium (e.g. from pypdfium2); the native lib is
# never committed. Used only by examples/pdfium_compare.rs.
pdfium-render = "0.8"

[package.metadata.docs.rs]
all-features = true
rustdoc-args = ["--cfg", "docsrs"]