pyRegTab

RegTab: pattern-driven data extraction from document tables with regular structure — the Python port of jRegTab with a native Rust core.

pyRegTab compiles RTL (Regular Table Language) patterns into abstract table patterns (ATP), matches them against a table's syntactic layer (ITM), and interprets the match into a relational recordset:

TableSyntax → RtlCompiler/TablePattern → AtpMatcher → TableInterpreter → Recordset

pyRegTab 0.7.1 ≙ jRegTab 0.7.1 (same API, same semantics, same test corpus), including the embedded RTL DSL pyregtab.dsl — a port of jRegTab's ru.icc.regtab.dsl (added upstream in jRegTab 0.3.0). Python-side extras on top of the Java API: AtpMatcher.match_many (parallel batch matching), Recordset.to_pandas(), Recordset.to_csv(), RtlCompileError.line/.col attributes, CellDerivedItem.span (the item's byte range within the raw cell text), and the CLI runner python -m pyregtab.runner (see below).

Since 0.5.1 the language and the interpreter follow jRegTab 0.6.0–0.7.1: JOIN(K) is the record product (a cross product for K = ∅, an equi-join on the key K otherwise), the former folding operation is spelled CONCAT(K) — a pattern written for 0.5.x must replace JOIN(K) by CONCAT(K); the key K may name attributes (CONCAT(0, 1, 'A', 'B'), JOIN('Year')); skipped CONCAT/JOIN actions (a key mismatch, a shared attribute, a forgotten REC) are reported through TableInterpreter.diagnostics(); zero-width subrows match the empty sequence and {0}/{1} are accepted; cell-derived providers use a spatial index, so interpretation is linear in the number of cells (a 1.19 M-cell table interprets in seconds).

Installation

pip install pyregtab

Binary wheels are published for Windows, Linux and macOS (x86-64 / arm64), CPython ≥ 3.10 (one abi3 wheel per platform). Building from the sdist requires a Rust toolchain.

Example

from pyregtab import TableSyntax, RtlCompiler, AtpMatcher, TableInterpreter

syntax = TableSyntax(3, 3)
syntax.cell(0, 1).set_text("CA");  syntax.cell(0, 2).set_text("HU")
syntax.cell(1, 0).set_text("IKT"); syntax.cell(1, 1).set_text("0 Jan"); syntax.cell(1, 2).set_text("8 Feb")
syntax.cell(2, 0).set_text("SVO"); syntax.cell(2, 1).set_text("31 Jan"); syntax.cell(2, 2).set_text("40 Feb")

pattern = RtlCompiler.compile("""
    [ [] [VAL : 'AIRLINE'->AVP]+ ]
    [ [VAL : 'AIRPORT'->AVP]
      [VAL : (COL, ROW, CL)->REC, 'ND'->AVP " " VAL : 'MON'->AVP]+ ]+
""")

itm = AtpMatcher.match(pattern, syntax)     # InterpretableTable | None
rs = TableInterpreter().interpret(itm)      # Recordset
rs.schema.attributes                        # ['ND', 'AIRLINE', 'AIRPORT', 'MON']
rs[0]["ND"]                                 # '0'
df = rs.to_pandas()                         # extras: pip install pyregtab[pandas]

Patterns can also be built without RTL, via the fluent spec API (TablePattern.of(SubtablePattern.of(...)) — same factories as in Java, snake_case method names), and serialized back to RTL with AtpToRtlSerializer.serialize(pattern).

For a terser, RTL-like way to build patterns in code, use the embedded RTL DSL (pyregtab.dsl) — see Embedded RTL below.

Named Python predicates are attached to RTL via EXT('name'):

from pyregtab import Bindings

p = RtlCompiler.compile(
    "{ [ [EXT('isTotal') ? VAL : ST*->REC] []+ ] }+",
    Bindings.of().cell("isTotal", lambda cell: cell.text.startswith("Total")),
)

Embedded RTL

The pyregtab.dsl module is a fluent DSL that reads almost like RTL but is ordinary Python — with IDE completion, structural typing, pattern composition via plain variables, and Python callables as escape-hatch constraints. It builds the same TablePattern objects as the compiler (verified byte-for-byte against RtlCompiler.compile for a representative set of tasks in tests/test_dsl.py).

from pyregtab.dsl import *

# RTL: { [ [VAL : ST*->REC] [VAL]{2} []+ ]
#        [ []               [VAL]{4} []+ ] }+
p = table(
    subtable(
        row(cell(VAL, rec(ST.unbounded())), cell(VAL).exactly(2), skip().one_or_more()),
        row(skip(),                         cell(VAL).exactly(4), skip().one_or_more()),
    ).one_or_more())

Method names are snake_case (.one_or_more(), .and_(), .split_by()); the vocabulary constants (VAL, ST, COL, C(n), …) match RTL. See the Embedded RTL guide for the full mapping and the where(...) escape hatch.

API mapping (Java → Python)

Architecture

Everything after the Python call boundary runs in a native core written in Rust (pyregtab._core, built with PyO3 and maturin); the Python layer is a thin re-export.

• grammar/RTL.g4 — the normative specification of the RTL language (a verbatim copy from jRegTab; the upstream commit and the grammar's SHA-256 are recorded in grammar/UPSTREAM). The core's parser is a hand-written lexer + recursive descent that structurally follows the grammar rules. A CI job (tools/check_grammar_sync.py) fails the build if the copy drifts from the pinned hash, and — when a jRegTab read token is available — cross-checks it byte-for-byte against the upstream commit.
• conformance/ — the shared RTL conformance corpus (also pinned from jRegTab, see conformance/UPSTREAM and conformance/README.md). Both implementations must compile every positive case to the same canonical form and reject every negative case; the corpus runs in CI of both projects. Any RTL language change flows: RTL.g4 in jregtab → corpus extension → both parsers → green corpus in both CIs.
• Regular expressions in RTL constraints are executed by the Rust regex crate (linear-time). The reference fixture corpus uses no lookaround/backreferences (audited), so the dialect is compatible with java.util.regex on this corpus. Documented divergences from Java: \d/\s/\w are Unicode-aware in regex (ASCII in Java), and SUBSTR indices count code points (UTF-16 units in Java) — identical behavior on the entire reference corpus.

Testing

pytest tests runs (1 981 tests):

• the full benchmark suite — tasks 001–150 (Foofah, RegTab, Baikal), every fixture variant, both via RTL patterns and via ATP patterns built with the Python spec API (1 500 task variants in total; fixtures are copied verbatim from jRegTab into tests/fixtures/tasks, ATP builders are mechanically translated from the Java tests by tools/translate_atp.py);
• embedded RTL DSL parity — 26 representative tasks/constructs built with pyregtab.dsl produce byte-identical ATP to RtlCompiler.compile (tests/test_dsl.py);
• the RTL conformance corpus (positive canonical forms, fixed points, negative rejections);
• RTL↔ATP round-trip for tasks 001–050;
• API unit tests (syntax layer, extractors, EXT bindings, custom predicates, transformations, interpreter options, GIL-released batch matching from a thread pool and via AtpMatcher.match_many);
• CONCAT/JOIN semantics, named keys, diagnostics, strict preconditions, zero-width subrows and {0}/{1} (tests/test_join_concat.py, a port of jRegTab's TableInterpreterMultiRecordTest).

cargo test additionally runs the conformance corpus, an end-to-end smoke test against the native core alone, the working-state unit tests for CONCAT/JOIN (ports of WorkingStateConcatTest/WorkingStateJoinTest), the zero-width matcher cases, and a randomized equivalence test of the indexed provider against the reference definition Υ^{J,k}_{τ,κ} (full scan + sort). Differential testing against the Java reference (tools/differential.py + tools/RecordsetDumpMain.java) compares recordsets cell-by-cell on all 750 task variants; the 244 ATBench solutions of regtab-eval-on-atbench are byte-identical between the jRegTab 0.7.1 runner and python -m pyregtab.runner.

IDE support

Install Regular Table Language (RTL) from the VS Code Marketplace (ext install regtab.regtab): syntax highlighting for .rtl files and for RTL embedded in Python strings passed to RtlCompiler.compile(...), plus compile diagnostics and a live match preview against CSV fixtures. The extension sources are at regtab/vscode-rtl; a TextMate bundle for IntelliJ/PyCharm and other TextMate editors is under ide/.

RTL is also validated at compile time: RtlCompiler.compile(...) raises RtlCompileError with a line:col position on an invalid pattern.

Command-line runner

python -m pyregtab.runner <solution.rtl> (or python -m pyregtab <solution.rtl>) reads ./input.csv, applies the pattern (RtlCompiler.compile → AtpMatcher.match → TableInterpreter with RECORD_FIRST → pattern.transform) and writes ./output.csv — the schema as the header row, every field double-quoted, a missing value as the empty string, UTF-8, LF. It is the drop-in counterpart of the regtab-runner jar used by regtab-eval-on-atbench: same I/O contract, same exit codes (0 — success, 1 — the pattern did not match, 2 — invalid usage), same CSV parsing (RFC 4180 with multi-line quoted cells, ragged rows padded); on the 244 ATBench solutions of that project the two runners produce byte-identical output.csv. --input/--output override the file names, --strict turns interpreter diagnostics into errors.

Development

python -m venv .venv && . .venv/bin/activate   # or .venv\Scripts\activate
pip install maturin pytest
maturin develop --release
pytest tests -q

License

MIT