endomorphosis's picture
Incremental country-laws-ir GraphRAG release for justicedao/ipfs_libya_laws_ir
1c25384 verified
Raw History Blame Contribute Delete
7.19 kB
"""Reconstruct empty parent law bodies from article children.
Never invents legal text. Glue uses the same normalized strings as child
corpus fields. Eligible empty parents always get a law row (full glue,
streamed prefix, or title/number stub) — they are not dropped.
"""
from __future__ import annotations
from dataclasses import dataclass, field
import os
import re
import unicodedata
from typing import Any, Iterable
from .structure import normalize_legal_text
RECONSTRUCT_MAX_PARENT_CHARS = 2_000_000
RECONSTRUCT_STUB_MAX_CHARS = 8_192
RECONSTRUCT_PROCESS_BUDGET_BYTES = 512 * 1024 * 1024
RSS_ABORT_SKIP_FULL_CONCAT = frozenset({"france", "eu", "denmark"})
PARENT_TRUNCATED_MAX_CHARS = 40
_DIGIT_FOLD = str.maketrans(
{
"٠": "0",
"١": "1",
"٢": "2",
"٣": "3",
"٤": "4",
"٥": "5",
"٦": "6",
"٧": "7",
"٨": "8",
"٩": "9",
"۰": "0",
"۱": "1",
"۲": "2",
"۳": "3",
"۴": "4",
"۵": "5",
"۶": "6",
"۷": "7",
"۸": "8",
"۹": "9",
"〇": "0",
"零": "0",
}
)
_NUM_RUN = re.compile(r"\d+|\D+")
def fold_digits(value: str) -> str:
return unicodedata.normalize("NFKC", value or "").translate(_DIGIT_FOLD)
def article_sort_key(article_number: str, source_id: str) -> tuple:
folded = fold_digits(article_number or "")
nums: list[int] = []
rest: list[str] = []
for run in _NUM_RUN.findall(folded) if folded else []:
if run.isdigit():
nums.append(int(run))
else:
rest.append(run)
if not folded:
return (0, (), "", source_id or "")
if nums:
return (1, tuple(nums), "".join(rest), source_id or "")
return (2, (), folded, source_id or "")
def glue_piece(title: str, article_number: str, body: str) -> str:
"""One child's contribution to a reconstructed parent (normalized strings)."""
title = title or ""
article_number = article_number or ""
body = body or ""
chunk = title if title else article_number
if chunk and not body.lstrip().startswith(chunk):
return f"{chunk}\n{body}" if body else chunk
return body
def glue_parent(pieces: Iterable[str]) -> str:
return "\n\n".join(p for p in pieces if p)
@dataclass
class ReconstructResult:
body: str
reconstructed_from_articles: bool
reconstruction_truncated: bool
reconstruction_gap_note: str
reconstruction_article_count: int
reconstruction_article_ids: list[str] = field(default_factory=list)
reconstruction_article_sha256: list[str] = field(default_factory=list)
extra_bytes: int = 0
is_stub: bool = False
def extra_fields(self) -> dict[str, Any]:
return {
"reconstructed_from_articles": self.reconstructed_from_articles,
"reconstruction_truncated": self.reconstruction_truncated,
"reconstruction_gap_note": self.reconstruction_gap_note,
"reconstruction_article_count": self.reconstruction_article_count,
"reconstruction_article_ids": list(self.reconstruction_article_ids),
"reconstruction_article_sha256": list(self.reconstruction_article_sha256),
}
def parent_needs_reconstruct(parent_body: str, n_articles: int) -> bool:
if n_articles < 1:
return False
body = parent_body or ""
return (not body) or len(body) < PARENT_TRUNCATED_MAX_CHARS
def reconstruct_on() -> bool:
return os.environ.get("COUNTRY_LAWS_RECONSTRUCT", "1").strip() not in {"0", "false", "off"}
def force_full_concat() -> bool:
return os.environ.get("COUNTRY_LAWS_RECONSTRUCT_FULL", "").strip() in {"1", "true", "on"}
def process_budget_bytes() -> int:
raw = os.environ.get("RECONSTRUCT_PROCESS_BUDGET_BYTES")
if raw and raw.isdigit():
return int(raw)
return RECONSTRUCT_PROCESS_BUDGET_BYTES
def reconstruct_parent(
*,
slug: str,
parent_body: str,
parent_title: str,
children: list[dict[str, str]],
running_extra_bytes: int,
sha256_hex,
) -> ReconstructResult:
"""Apply the five-way outcome table. ``children`` already normalized."""
ordered = sorted(
children,
key=lambda c: article_sort_key(c.get("article_number") or "", c.get("id") or ""),
)
ids = [c.get("id") or "" for c in ordered]
shas = [sha256_hex((c.get("body") or "").encode("utf-8")) for c in ordered]
n = len(ordered)
if n < 1 or not parent_needs_reconstruct(parent_body, n):
return ReconstructResult(
body=parent_body,
reconstructed_from_articles=False,
reconstruction_truncated=False,
reconstruction_gap_note="",
reconstruction_article_count=0,
)
stub_note = ""
if slug in RSS_ABORT_SKIP_FULL_CONCAT and not force_full_concat():
stub_note = "rss_abort_slug"
elif running_extra_bytes >= process_budget_bytes():
stub_note = "reconstruct_budget"
if stub_note:
parts: list[str] = []
if parent_title:
parts.append(parent_title)
for child in ordered:
label = child.get("title") or child.get("article_number") or ""
if label:
parts.append(label)
stub = "\n".join(parts)[:RECONSTRUCT_STUB_MAX_CHARS]
return ReconstructResult(
body=stub,
reconstructed_from_articles=True,
reconstruction_truncated=False,
reconstruction_gap_note=stub_note,
reconstruction_article_count=n,
reconstruction_article_ids=ids,
reconstruction_article_sha256=shas,
extra_bytes=0,
is_stub=True,
)
acc: list[str] = []
size = 0
truncated = False
for child in ordered:
piece = glue_piece(
child.get("title") or "",
child.get("article_number") or "",
child.get("body") or "",
)
if not piece:
continue
extra = (2 if acc else 0) + len(piece)
if size + extra > RECONSTRUCT_MAX_PARENT_CHARS:
room = RECONSTRUCT_MAX_PARENT_CHARS - size - (2 if acc else 0)
if room > 0:
acc.append(piece[:room])
size = RECONSTRUCT_MAX_PARENT_CHARS
truncated = True
break
acc.append(piece)
size += extra
body = glue_parent(acc)
return ReconstructResult(
body=body,
reconstructed_from_articles=True,
reconstruction_truncated=truncated,
reconstruction_gap_note="",
reconstruction_article_count=n,
reconstruction_article_ids=ids,
reconstruction_article_sha256=shas,
extra_bytes=len(body.encode("utf-8")),
is_stub=False,
)
def expected_glue_from_children(children: list[dict[str, str]]) -> str:
ordered = sorted(
children,
key=lambda c: article_sort_key(c.get("article_number") or "", c.get("id") or ""),
)
return glue_parent(
glue_piece(c.get("title") or "", c.get("article_number") or "", c.get("body") or "")
for c in ordered
)