epub-blocks is a dependency-free Python library for reproducibly extracting
ordered, structured text blocks from EPUB 2 and EPUB 3 files. A declarative
recipe selects an edition’s XHTML, assigns block types, and generates stable
identifiers without edition-specific Python code or a pre-existing output
table.
It requires Python 3.13 or later. The project is pre-1.0, so its public API and recipe format may still make breaking changes.
Version 0.10.0 adds native inline errata: explicit,
block-scoped literal corrections after plain-text extraction. Existing recipes
without errata keep their extraction behavior and compiled digests, and recipe
format remains "1". Statement and branch test coverage remain at 100%.
See the changelog.
python -m pip install epub-blocksThe core runtime package has no third-party dependencies.
An optional Textual interface now helps select documents, define section codes, classify block patterns, and preview a draft recipe before saving. It uses no LLM and generates an ordinary version-1 recipe. Unicode inline markup is the default. Install the optional interface with:
python -m pip install "epub-blocks[wizard]"
epub-blocks-wizard book.epub draft.recipe.jsonSee the wizard guide for save/resume, validation, and the structural cases that still need manual recipe editing.
The package:
- safely reads EPUB package metadata and XHTML spine documents;
- selects source documents and blocks with case-insensitive globs;
- normalizes extracted text and omits configured EPUB semantic elements;
- derives groups from source paths, heading markers, or explicit transitions;
- generates flat and nested identifiers with per-group counters or source paths;
- assigns output types through ordered source-block rules;
- handles joins, splits, omissions, skipped blocks, and inserted material;
- selects whole compound blocks and intentional empty structural blocks;
- inserts recipe-selected whitespace before/after nested elements without splitting ordinary inline word wrappers;
- preserves recipe-selected spans and milestones as XML fragments or configurable Unicode delimiters, without requiring new paragraph numbers;
- pins both the EPUB and the compiled extraction plan with SHA-256; and
- writes headerless
id,type,textTSV records.
A recipe is sufficient to generate a new base text. It does not need a
separately stored list of expected identifiers or types. Existing outputs may
be useful as test oracles while authoring a recipe, but they are not recipe
inputs and are not read by epub-blocks.
The package does not compare editions, select preferred readings, create witness manifests, or emit stand-off annotations. It can generate a marked-up TSV for a separate converter to derive plain text and stand-off together.
New in 0.10.0: recipes can also declare short, block-scoped literal errata for plain-text output. Each correction must match exactly once after extraction; the library never infers editorial choices. Errata require epub-blocks 0.10.0 or later.
Use the lower-level API to discover source blocks before writing a recipe:
from epub_blocks import extract_blocks
for block in extract_blocks(
"book.epub",
include_documents=["*chapter*.xhtml"],
exclude_classes=["image-caption"],
):
print(block.source_locator, block.tag, sorted(block.classes), block.text)A locator such as s008:text/chapter-01.xhtml#1.3.2 combines the one-based
spine position, EPUB-internal document path, and element path within the XHTML
body. Recipes use the EPUB-local portion: text/chapter-01.xhtml#1.3.2.
Recipe version 1 remains the serialized format in epub-blocks 0.10.0. TSV serialization uses literal fields, instead of CSV-style quoting:
{
"recipe_version": "1",
"metadata": {"name": "Example edition"},
"epub": {
"identifier": "9780000000000",
"sha256": "f4f9c2d902a41b80732b2dce7ad01a57f615859c21417a5019e3cd8e4d271282"
},
"normalization": {
"collapse_whitespace": true,
"strip": true,
"unicode_normalization": "NFC"
},
"omit_epub_types": ["noteref", "pagebreak"],
"source_blocks": {
"include_documents": ["text/chapter-*.xhtml"],
"exclude_classes": ["image-caption"]
},
"output": {
"groups": {
"source_marker": {
"pattern": "^CHAPTER\\s+([IVXLCDM]+)\\b",
"case_insensitive": true
},
"capture_kind": "roman",
"capture_width": 2
},
"identifiers": {
"block": {
"template": "{group}.{number:03d}",
"start": 1
},
"line": {
"template": "{group}.{block:03d}.{number:02d}",
"start": 1
}
},
"default": {"type": "paragraph", "role": "block"},
"rules": [
{
"match": {"tag": "h1"},
"type": "heading",
"role": "fixed",
"id": "{group}.000",
"remove_prefix": {
"pattern": "^CHAPTER\\s+[IVXLCDM]+\\s+",
"case_insensitive": true
}
},
{
"match": {"classes": ["verse-first"]},
"type": "line",
"role": "line-start"
},
{
"match": {"classes": ["verse"]},
"type": "line",
"role": "line"
}
],
"skip_source": [],
"replacements": [],
"insertions": [],
"compiled_sha256": "28354978484f35525c616d138bdeb8f801039d665235ac88bd02f1a25791c410"
}
}This recipe generates identifiers such as 01.000, 01.001, 01.002.01,
and 01.002.02. Rules are matched against source structure; no expected text
or identifier table is consulted.
See the complete recipe specification and packaged JSON Schema.
The recipe's text.markup object chooses "format": "xml" or
"format": "delimiters". Both serializers use the same structural rules,
references, normalized text, and annotation boundaries. For example:
XML: A <em>quiet</em> word.<page label="iv"/>
Delimiters: A ⧼quiet⧽ word.⟦iv⟧
Milestones between text blocks can attach to the following block without consuming a number. Trailing milestones require an explicit policy. Literal text is escaped so it cannot be mistaken for markup. See the markup contract for the complete configuration, grammar, escaping rules, normalization, slicing, and attachment semantics.
The third format, "literal", supports established notations such as
"paragraph": ["¶", "¶"] or "line-break": ["∥", ""] without an escaping
layer. It uses the same extraction policy, but source text may collide with
markup tokens: literal output is not generally reversible. Use XML or
escaped delimiters for unambiguous downstream decoding. Literal mode also
leaves raw control characters untouched; the TSV writer still rejects TAB,
CR and LF in fields.
A recipe can distinguish layout spans from ordinary word wrappers:
"text": {
"block_boundaries": {
"tags": ["p", "li"],
"separator": " ",
"rules": [
{"match": {"tag": "span", "classes_all": ["space"]}, "before": " "}
]
}
}This inserts a space before a retained nested span.space, including an
empty one, without discarding its contents. Add after for a trailing space.
Other spans remain transparent. Rules are first-match-wins and override the
tag shorthand for a matching element. Whitespace enters the character stream
before normalization and slicing; optional markup can record its source
without losing the space when markup is removed. See the
boundary specification and
spacing/markup contract.
epub-blocks book.epub recipe.json records.tsvThe output is headerless TSV with id, type, and extracted text columns.
Fields are separated by literal tabs and records by LF. Quotes and backslashes
are written unchanged: there is no CSV quoting, quote doubling, or TSV escaping.
Embedded TAB, CR, or LF characters in any field are rejected rather than
silently altering text or producing ambiguous rows. Normalize text or encode
such characters in the markup before writing. A failed write leaves any
existing output file intact.
from epub_blocks import extract_recipe_file
records = extract_recipe_file("book.epub", "recipe.json")
for record in records:
print(record.block_id, record.block_type, record.text)compile_recipe_file returns the immutable in-memory extraction plan.
For recipe authoring, extract_recipe_candidates(epub_path, recipe) returns
source blocks with locators, tags, classes, and plain normalized text, using
the recipe's source-element and text policies. It checks the EPUB hash and
identifier policy but does not apply output rules or require/verify the
compiled-plan digest. A syntactically valid draft output section is still required; use
load_recipe to read JSON. Final output references, coverage, and detached
milestone placement are checked by compilation, not candidate inspection.
Since 0.6.0, a recipe may omit epub.identifier only for an EPUB with no
nonempty package identifiers; the archive SHA-256 is still mandatory. A
supplied mismatched identifier never falls back to hash-only checking.
The compiled plan's epub_identifier is None in this case. See the
source-pin contract.
Fragment and extract_fragments provide lower-level access to selected XHTML
subtrees. The supported import surface is the names exported by epub_blocks.
EPUB files are untrusted ZIP and XML input. Default APIs bound archive size,
individual and cumulative reads, compression ratios, XML size, element count,
and nesting depth. Duplicate or unsafe paths and encrypted members are rejected.
External DTDs are never loaded; nonempty internal subsets, entity declarations, other
doctypes, and unknown named character references are rejected.
Version 0.7.0 additionally accepts an empty internal subset ([]),
including XML whitespace between the brackets. The declaration is stripped,
not interpreted; no custom entities or DTD rules are enabled.
Recipes are trusted local configuration because they contain regular
expressions. epub-blocks never executes EPUB scripts or fetches resources from
the network.
uv sync
uv run ruff format --check .
uv run ruff check .
uv run pyright
uv run coverage erase
uv run coverage run -m unittest discover -s tests
uv run coverage report
uv run python -m build
uv run twine check --strict dist/*
uv run pyright --verifytypes epub_blocks --ignoreexternalCoverage includes branches and enforces a 100% minimum. Tests build synthetic EPUBs; no EPUB files are committed. See CONTRIBUTING.md and CHANGELOG.md.
epub-blocks is available under the
MIT License.