src/clean_samorzad_gov_pl.py

#!/usr/bin/env python3
"""Convert dawidmajewski/samorzad-gov-pl-articles to DynaWord.

The input is the published source dataset. Institution names are frozen in a
committed publisher manifest generated from each tenant homepage's
``og:site_name`` metadata. The output uses DynaWord's canonical eight columns
and writes a URL-bearing attribution sidecar for CC-BY-SA compliance.
"""
from __future__ import annotations

import argparse
import hashlib
import html as html_module
import http.client
import json
import os
import re
import shutil
import tempfile
import time
import urllib.request
from pathlib import Path

import pyarrow as pa
import pyarrow.dataset as ds
import pyarrow.parquet as pq

SOURCE = "samorzad_gov_pl"
LICENSE = "CC-BY-SA-4.0"
LICENSE_URL = "https://creativecommons.org/licenses/by-sa/4.0/"
UPSTREAM_DATASET = "dawidmajewski/samorzad-gov-pl-articles"
UPSTREAM_REVISION = "d5eaefc32c3f17cfa5a01d14573f8e4ee7d43385"
UPSTREAM_TEXT_SHA256 = "6c35e1dc0bd35863ddb2f4e43d5bd08264c1fc27affdd60684ab5ecbc5457f63"
UPSTREAM_MANIFEST_SHA256 = "97b5a266b2fd4f043648b7d3c757721cd826c7ca1048301761084b0abdee5c48"
UPSTREAM_EXPECTED = {
    "manifest_sha256": UPSTREAM_MANIFEST_SHA256,
    "dataset_id": UPSTREAM_DATASET,
    "schema_version": "1.0.0",
    "record_count": 81_422,
    "record_text_sha256": UPSTREAM_TEXT_SHA256,
}
MIN_CHARS = 200
MIN_POLISH_RATIO = 0.005
POLISH_RE = re.compile(r"[ąćęłńóśźżĄĆĘŁŃÓŚŹŻ]")
ALPHA_RE = re.compile(r"[^\W\d_]", re.UNICODE)
_SITE_NAME_RE = re.compile(
    r"<meta\b(?=[^>]*\bproperty\s*=\s*['\"]og:site_name['\"])(?=[^>]*\bcontent\s*=\s*(['\"])(.*?)\1)[^>]*>",
    re.IGNORECASE | re.DOTALL,
)
_TITLE_RE = re.compile(r"<title\b[^>]*>(.*?)</title>", re.IGNORECASE | re.DOTALL)
DYNAWORD_SCHEMA = pa.schema(
    [
        ("id", pa.string()),
        ("text", pa.string()),
        ("source", pa.string()),
        ("added", pa.string()),
        ("created", pa.string()),
        ("token_count", pa.int64()),
        ("license", pa.string()),
        ("author", pa.string()),
    ]
)


def _clean_html_text(value: str) -> str:
    value = re.sub(r"<[^>]+>", " ", value)
    return " ".join(html_module.unescape(value).split())


def publisher_name_from_html(html: str) -> str:
    """Extract the human-readable institution name from a tenant homepage."""
    match = _SITE_NAME_RE.search(html)
    if match:
        return _clean_html_text(match.group(2))
    match = _TITLE_RE.search(html)
    if not match:
        return ""
    title = _clean_html_text(match.group(1))
    prefix, suffix = "Strona główna - ", " - Portal gov.pl"
    if title.startswith(prefix) and title.endswith(suffix):
        return title[len(prefix) : -len(suffix)].strip()
    return ""


def _fetch(url: str, attempts: int = 5) -> str:
    last_error: Exception | None = None
    for attempt in range(attempts):
        try:
            request = urllib.request.Request(
                url,
                headers={"User-Agent": "Polish-DynaWord-source-builder/1.0"},
            )
            with urllib.request.urlopen(request, timeout=45) as response:
                return response.read().decode("utf-8", "replace")
        except (OSError, TimeoutError, http.client.HTTPException) as error:
            last_error = error
            time.sleep(2**attempt)
    raise RuntimeError(f"Could not fetch {url}: {last_error}")


def refresh_publishers(tenants: list[str], output: Path) -> dict[str, str]:
    """Fetch and freeze one verified display name for every source tenant."""
    publishers: dict[str, str] = {}
    failures: list[str] = []
    for index, tenant in enumerate(sorted(set(tenants)), 1):
        url = f"https://samorzad.gov.pl/web/{tenant}"
        try:
            name = publisher_name_from_html(_fetch(url))
        except RuntimeError:
            name = ""
        if not name:
            failures.append(tenant)
        else:
            publishers[tenant] = name
        print(f"publishers {index}/{len(set(tenants))}: {tenant} -> {name or 'FAILED'}")
    if failures:
        raise RuntimeError(
            "Publisher-name refresh incomplete; retry before release: "
            + ", ".join(failures)
        )
    output.parent.mkdir(parents=True, exist_ok=True)
    output.write_text(
        json.dumps(
            {
                "schema_version": "1.0.0",
                "source": "tenant homepage og:site_name",
                "publishers": publishers,
            },
            ensure_ascii=False,
            indent=2,
            sort_keys=True,
        )
        + "\n",
        encoding="utf-8",
    )
    return publishers


def load_publishers(path: Path) -> dict[str, str]:
    data = json.loads(path.read_text(encoding="utf-8"))
    publishers = data.get("publishers")
    if data.get("schema_version") != "1.0.0" or not isinstance(publishers, dict):
        raise ValueError(f"Invalid publisher manifest: {path}")
    if not all(isinstance(k, str) and isinstance(v, str) and v.strip() for k, v in publishers.items()):
        raise ValueError(f"Empty or invalid publisher in {path}")
    return publishers


def _file_sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(chunk)
    return digest.hexdigest()


def verify_upstream_release(
    input_path: Path,
    manifest_path: Path,
    expected: dict[str, object] = UPSTREAM_EXPECTED,
) -> dict:
    """Fail closed unless input exactly matches the pinned published release."""
    input_path = input_path.resolve()
    manifest_path = manifest_path.resolve()
    raw_manifest = manifest_path.read_bytes()
    if hashlib.sha256(raw_manifest).hexdigest() != expected["manifest_sha256"]:
        raise ValueError("Upstream manifest hash mismatch")
    manifest = json.loads(raw_manifest)
    for key in ("dataset_id", "schema_version", "record_count", "record_text_sha256"):
        if manifest.get(key) != expected[key]:
            raise ValueError(f"Upstream manifest {key} mismatch")

    artifact_paths: list[Path] = []
    artifact_records = 0
    for artifact in manifest.get("artifacts", []):
        path = (manifest_path.parent / artifact["path"]).resolve()
        if path.parent != input_path:
            raise ValueError(f"Upstream artifact outside input directory: {path}")
        if not path.is_file():
            raise ValueError(f"Missing upstream artifact: {path}")
        if path.stat().st_size != artifact["bytes"]:
            raise ValueError(f"Upstream artifact size mismatch: {path.name}")
        if _file_sha256(path) != artifact["sha256"]:
            raise ValueError(f"Upstream artifact hash mismatch: {path.name}")
        if pq.ParquetFile(path).metadata.num_rows != artifact["records"]:
            raise ValueError(f"Upstream artifact row-count mismatch: {path.name}")
        artifact_paths.append(path)
        artifact_records += artifact["records"]

    actual_paths = sorted(path.resolve() for path in input_path.glob("*.parquet"))
    if sorted(artifact_paths) != actual_paths:
        raise ValueError("Upstream artifact set differs from manifest")
    if artifact_records != expected["record_count"]:
        raise ValueError("Upstream artifact record total mismatch")

    source_dataset = ds.dataset([str(path) for path in actual_paths], format="parquet")
    if source_dataset.schema.names != manifest["record_fields"]:
        raise ValueError("Upstream Parquet schema differs from manifest")
    source_rows = source_dataset.to_table(columns=["record_id", "text"]).to_pylist()
    if len(source_rows) != expected["record_count"]:
        raise ValueError("Upstream dataset record count mismatch")
    source_rows.sort(key=lambda row: row["record_id"])
    if len({row["record_id"] for row in source_rows}) != len(source_rows):
        raise ValueError("Upstream record IDs are not unique")
    text_digest = hashlib.sha256()
    for row in source_rows:
        text_digest.update(row["record_id"].encode())
        text_digest.update(b"\0")
        text_digest.update(row["text"].encode())
        text_digest.update(b"\n")
    if text_digest.hexdigest() != expected["record_text_sha256"]:
        raise ValueError("Upstream record-text digest mismatch")
    return manifest


def normalize_record(
    source: dict,
    publishers: dict[str, str],
    token_count: int,
    added: str,
) -> tuple[dict, dict]:
    tenant = str(source["tenant"])
    try:
        publisher = publishers[tenant]
    except KeyError as error:
        raise KeyError(f"No publisher name for tenant {tenant}") from error
    record_id = f"{SOURCE}_{source['record_id']}"
    created = source.get("published_date") or ""
    source_url = str(source["source_url"])
    final_url = str(source.get("final_url") or source_url)
    row = {
        "id": record_id,
        "text": str(source["text"]),
        "source": SOURCE,
        "added": added,
        "created": str(created),
        "token_count": int(token_count),
        "license": LICENSE,
        "author": publisher,
    }
    attribution = {
        "id": record_id,
        "title": str(source.get("title") or ""),
        "source_url": source_url,
        "final_url": final_url,
        "attribution_url": final_url,
        "publisher": publisher,
        "tenant": tenant,
        "publisher_type": str(source.get("publisher_type") or ""),
        "license": LICENSE,
        "license_url": LICENSE_URL,
        "modified": True,
        "modification_notice": (
            "Article text was extracted from the source HTML. This DynaWord "
            "release normalizes metadata and may exclude short, non-Polish or "
            "duplicate records; the retained text is otherwise unchanged."
        ),
    }
    return row, attribution


def _polish_ratio(text: str) -> float:
    letters = ALPHA_RE.findall(text)
    return len(POLISH_RE.findall(text)) / len(letters) if letters else 0.0


def _text_hash(text: str) -> bytes:
    return hashlib.sha256(text.encode("utf-8")).digest()


def existing_text_hashes(repo_root: Path, output_path: Path) -> set[bytes]:
    """Load exact-text hashes from all existing DynaWord source Parquets."""
    hashes: set[bytes] = set()
    output_resolved = output_path.resolve()
    for path in sorted((repo_root / "data").glob("*/*.parquet")):
        if path.resolve() == output_resolved or path.parent.name == SOURCE:
            continue
        for batch in pq.ParquetFile(path).iter_batches(columns=["text"], batch_size=4096):
            hashes.update(_text_hash(text) for text in batch.column(0).to_pylist() if text)
    return hashes


def _fsync_path(path: Path) -> None:
    with path.open("rb") as handle:
        os.fsync(handle.fileno())


def _write_text_synced(path: Path, content: str) -> None:
    with path.open("w", encoding="utf-8") as handle:
        handle.write(content)
        handle.flush()
        os.fsync(handle.fileno())


def _validate_staged_release(staged_dir: Path) -> None:
    parquet_path = staged_dir / f"{SOURCE}.parquet"
    sidecar_path = staged_dir / f"{SOURCE}.attribution.jsonl"
    stats = json.loads((staged_dir / f"{SOURCE}.stats.json").read_text())
    table = pq.read_table(parquet_path)
    if table.schema != DYNAWORD_SCHEMA or table.num_rows != stats["kept"]:
        raise ValueError("Staged Parquet does not match canonical schema or stats")
    ids = table["id"].to_pylist()
    authors = table["author"].to_pylist()
    if len(ids) != len(set(ids)):
        raise ValueError("Staged IDs are not unique")
    sidecar_count = 0
    with sidecar_path.open(encoding="utf-8") as handle:
        for sidecar_count, (line, record_id, author) in enumerate(
            zip(handle, ids, authors, strict=True),
            1,
        ):
            attribution = json.loads(line)
            if attribution["id"] != record_id or attribution["publisher"] != author:
                raise ValueError("Staged sidecar does not match Parquet order")
            if attribution["attribution_url"] != attribution["final_url"]:
                raise ValueError("Staged attribution URL is not the fetched final URL")
    if sidecar_count != stats["kept"]:
        raise ValueError("Staged sidecar count does not match stats")


def _replace_release(staged_dir: Path, output_dir: Path) -> None:
    """Replace a validated artifact set, restoring the previous set on error."""
    names = (
        f"{SOURCE}.parquet",
        f"{SOURCE}.attribution.jsonl",
        f"{SOURCE}.stats.json",
    )
    backup_dir = staged_dir / "backup"
    backup_dir.mkdir()
    had_previous: set[str] = set()
    for name in names:
        destination = output_dir / name
        if destination.exists():
            shutil.copy2(destination, backup_dir / name)
            _fsync_path(backup_dir / name)
            had_previous.add(name)
    try:
        for name in names:
            os.replace(staged_dir / name, output_dir / name)
        directory_fd = os.open(output_dir, os.O_RDONLY)
        try:
            os.fsync(directory_fd)
        finally:
            os.close(directory_fd)
    except BaseException:
        for name in names:
            destination = output_dir / name
            backup = backup_dir / name
            if name in had_previous:
                os.replace(backup, destination)
            elif destination.exists():
                destination.unlink()
        raise


def build(
    input_path: Path,
    manifest_path: Path,
    repo_root: Path,
    publishers_path: Path,
    added: str,
    refresh: bool,
) -> dict:
    manifest = verify_upstream_release(input_path, manifest_path)
    source_dataset = ds.dataset(str(input_path), format="parquet")
    source_table = source_dataset.to_table(
        columns=[
            "record_id",
            "source_url",
            "final_url",
            "tenant",
            "publisher_type",
            "title",
            "text",
            "published_date",
        ]
    )
    records = sorted(source_table.to_pylist(), key=lambda row: row["record_id"])
    tenants = sorted({row["tenant"] for row in records})
    tenant_types = {row["tenant"]: row["publisher_type"] for row in records}
    publishers = (
        refresh_publishers(tenants, publishers_path)
        if refresh
        else load_publishers(publishers_path)
    )
    missing = sorted(set(tenants) - set(publishers))
    if missing:
        raise RuntimeError(f"Publisher manifest misses {len(missing)} tenants: {missing}")

    output_dir = repo_root / "data" / SOURCE
    output_dir.mkdir(parents=True, exist_ok=True)
    parquet_path = output_dir / f"{SOURCE}.parquet"
    prior_hashes = existing_text_hashes(repo_root, parquet_path)
    seen = set(prior_hashes)

    candidates: list[dict] = []
    stats = {
        "read": len(records),
        "kept": 0,
        "drop_short": 0,
        "drop_lang": 0,
        "drop_dup": 0,
        "drop_ocr": 0,
        "chars": 0,
        "tokens": 0,
        "licenses": {LICENSE: 0},
        "authors_with_value": 0,
        "publisher_count": len(publishers),
        "local_public_tenant_count": sum(
            publisher_type != "central" for publisher_type in tenant_types.values()
        ),
        "central_tenant_count": sum(
            publisher_type == "central" for publisher_type in tenant_types.values()
        ),
        "redirected_urls": 0,
        "cross_source_hashes_loaded": len(prior_hashes),
        "license": LICENSE,
        "upstream_dataset": UPSTREAM_DATASET,
        "upstream_revision": UPSTREAM_REVISION,
        "upstream_record_text_sha256": UPSTREAM_TEXT_SHA256,
        "upstream_manifest_sha256": hashlib.sha256(manifest_path.read_bytes()).hexdigest(),
        "upstream_artifact_count": len(manifest["artifacts"]),
    }
    for record in records:
        text = str(record.get("text") or "").strip()
        if len(text) < MIN_CHARS:
            stats["drop_short"] += 1
            continue
        if _polish_ratio(text) < MIN_POLISH_RATIO:
            stats["drop_lang"] += 1
            continue
        digest = _text_hash(text)
        if digest in seen:
            stats["drop_dup"] += 1
            continue
        seen.add(digest)
        record["text"] = text
        candidates.append(record)

    import tiktoken

    encoder = tiktoken.get_encoding("cl100k_base")
    token_counts = [
        len(tokens)
        for tokens in encoder.encode_ordinary_batch(
            [record["text"] for record in candidates], num_threads=8
        )
    ]
    rows: list[dict] = []
    attribution_rows: list[dict] = []
    for source_row, token_count in zip(candidates, token_counts, strict=True):
        row, attribution = normalize_record(source_row, publishers, token_count, added)
        rows.append(row)
        attribution_rows.append(attribution)
        stats["kept"] += 1
        stats["chars"] += len(row["text"])
        stats["tokens"] += token_count
        stats["licenses"][LICENSE] += 1
        stats["authors_with_value"] += int(bool(row["author"]))
        stats["redirected_urls"] += int(
            attribution["source_url"] != attribution["final_url"]
        )

    table = pa.Table.from_pylist(rows, schema=DYNAWORD_SCHEMA)
    with tempfile.TemporaryDirectory(prefix=f".{SOURCE}-", dir=output_dir) as temp:
        staged_dir = Path(temp)
        staged_parquet = staged_dir / f"{SOURCE}.parquet"
        staged_sidecar = staged_dir / f"{SOURCE}.attribution.jsonl"
        staged_stats = staged_dir / f"{SOURCE}.stats.json"
        pq.write_table(table, staged_parquet, compression="zstd", row_group_size=4096)
        _fsync_path(staged_parquet)
        _write_text_synced(
            staged_sidecar,
            "".join(
                json.dumps(row, ensure_ascii=False, sort_keys=True) + "\n"
                for row in attribution_rows
            ),
        )
        _write_text_synced(
            staged_stats,
            json.dumps(stats, ensure_ascii=False, indent=2, sort_keys=True) + "\n",
        )
        _validate_staged_release(staged_dir)
        _replace_release(staged_dir, output_dir)
    return stats


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--input", type=Path, required=True, help="directory containing source Parquet shards")
    parser.add_argument(
        "--manifest",
        type=Path,
        help="pinned source manifest (default: MANIFEST next to the input directory)",
    )
    parser.add_argument("--repo-root", type=Path, default=Path(__file__).resolve().parents[1])
    parser.add_argument("--publishers", type=Path, default=Path(__file__).with_name("samorzad_gov_pl_publishers.json"))
    parser.add_argument("--added", required=True)
    parser.add_argument("--refresh-publishers", action="store_true")
    args = parser.parse_args()
    stats = build(
        input_path=args.input,
        manifest_path=args.manifest or args.input.parent / "manifest.json",
        repo_root=args.repo_root.resolve(),
        publishers_path=args.publishers,
        added=args.added,
        refresh=args.refresh_publishers,
    )
    print(json.dumps(stats, ensure_ascii=False, indent=2))


if __name__ == "__main__":
    main()