legex/legora.py

"""Ingest Legora tabular-review extractions from a single xlsx into per-country JSONL.
"""

import argparse
import html
import json
import logging
import re
import sys
from collections import defaultdict
from csv import DictReader
from datetime import date, datetime
from pathlib import Path

from openpyxl import load_workbook

from legex.config import settings
from legex.harvey import _gold_case_id_index
from legex.inference import _output_columns
from legex.utils import inference_path, norm_case_id, write_inference_jsonl

log = logging.getLogger(__name__)

LEGORA_FIELDS: tuple[str, ...] = (
    "case_id",
    "legal_subject_judgement",
    "trial_start_date",
    "trial_end_date",
    "dispute_value_nominal",
    "plaintiff_loosing_share",
    "court_cost_awarded_nominal",
    "party_compensation_awarded_nominal",
    "plaintiffs_all_count",
    "defendants_all_count",
    "plaintiff_no1_ISIC1_industry_category",
    "defendant_no1_ISIC1_industry_category",
)

# n-th occurrence of the field headers gives run name
GROUP_MODELS: dict[int, str] = {1: "legora-1", 2: "legora-2"}

CONFLICTS_PATH = Path("data/analysis/quality/legora_duplicate_conflicts.jsonl")
_EMPTY_LITERALS = {"", "—"}
_DATE_IN_NAME_RE = re.compile(r"(\d{4}-\d{2}-\d{2})")


def loose_name(s: str) -> str:
    """Filename key tolerant of separator mangling: unescape HTML entities,
    lowercase, collapse every non-alphanumeric run to one underscore."""
    return re.sub(r"[^0-9a-zà-￿]+", "_", html.unescape(str(s)).lower()).strip("_")


def tight_name(s: str) -> str:
    """Last-resort filename key: drop every non-alphanumeric character, so
    names differing only in punctuation placement compare equal."""
    return re.sub(r"[^0-9a-zà-￿]", "", html.unescape(str(s)).lower())


def load_manifest(csv_path: Path) -> dict[str, dict[str, tuple[str, str]]]:
    """Bundle manifest -> {"exact"|"loose"|"tight": {key: (cc, case_id)}}.
    """
    lookups: dict[str, dict[str, tuple[str, str]]] = {"exact": {}, "loose": {}, "tight": {}}
    ambiguous: dict[str, set[str]] = {"exact": set(), "loose": set(), "tight": set()}
    with open(csv_path, encoding="utf-8") as f:
        for rec in DictReader(f):
            target = (rec["cc"], rec["case_id"])
            for pass_, key in (
                ("exact", rec["file"]),
                ("loose", loose_name(rec["file"])),
                ("tight", tight_name(rec["file"])),
            ):
                table = lookups[pass_]
                if key in table and table[key] != target:
                    ambiguous[pass_].add(key)
                table.setdefault(key, target)
    for pass_, keys in ambiguous.items():
        for key in keys:
            del lookups[pass_][key]
        if keys:
            log.warning(f"manifest: dropped {len(keys)} ambiguous {pass_} key(s)")
    return lookups


def match_document(name: str, lookups: dict[str, dict[str, tuple[str, str]]]) -> tuple[str, str] | None:
    """(cc, case_id) for an exported document name, or None (junk/unknown)."""
    return (
        lookups["exact"].get(str(name))
        or lookups["loose"].get(loose_name(name))
        or lookups["tight"].get(tight_name(name))
    )


def _group_value_columns(header: list[str]) -> dict[int, dict[str, int]]:
    """{group -> {field -> column index}} from the header row.
    """
    occurrences: dict[str, list[int]] = defaultdict(list)
    for idx, cell in enumerate(header):
        name = str(cell).strip() if cell is not None else ""
        if name in LEGORA_FIELDS:
            occurrences[name].append(idx)
    counts = {f: len(occurrences[f]) for f in LEGORA_FIELDS}
    n_groups = min(counts.values())
    if n_groups < 1:
        missing = [f for f, n in counts.items() if n == 0]
        raise ValueError(f"export header is missing field column(s): {missing}")
    if len(set(counts.values())) != 1:
        raise ValueError(f"unbalanced field-column groups: {counts}")
    return {
        g: {f: occurrences[f][g - 1] for f in LEGORA_FIELDS}
        for g in range(1, n_groups + 1)
    }


def _clean(value: object) -> str:
    """Format-level canonicalisation of one cell (no value-level cleaning)."""
    if value is None:
        return ""
    if isinstance(value, datetime):  # Excel date cells arrive as datetimes
        return value.date().isoformat()
    if isinstance(value, date):
        return value.isoformat()
    if isinstance(value, float) and value.is_integer():
        return str(int(value))
    s = str(value).strip()
    if s in _EMPTY_LITERALS:
        return ""
    if s.lower() == "nonpecuniary":
        return "nonpecuniary"
    return s


def _infer_date_from_name(xlsx: Path) -> str | None:
    m = _DATE_IN_NAME_RE.search(xlsx.name)
    return m.group(1) if m else None


def ingest(
    xlsx: Path,
    manifest_csv: Path,
    prompt_version: str = "v3",
    source: str = "full_text",
    inference_date: str | None = None,
    conflicts_out: Path = CONFLICTS_PATH,
) -> None:
    inference_date = inference_date or _infer_date_from_name(xlsx)
    if not inference_date:
        raise ValueError(f"cannot derive inference date from {xlsx.name}; pass --inference_date")

    columns = _output_columns()
    columns.insert(columns.index("model") + 1, "inference_date")

    lookups = load_manifest(manifest_csv)
    wb = load_workbook(xlsx, read_only=True, data_only=True)
    if "Sheet1" not in wb.sheetnames:
        raise ValueError(f"{xlsx} missing Sheet1 (found {wb.sheetnames})")
    ws = wb["Sheet1"]
    rows_iter = ws.iter_rows(values_only=True)
    header = [c for c in next(rows_iter)]
    group_cols = _group_value_columns(header)

    # by_cc[model][cc] -> case_id -> {field: cleaned value}; export order kept.
    by_cc: dict[str, dict[str, dict[str, dict[str, str]]]] = {
        model: defaultdict(dict) for model in GROUP_MODELS.values()
    }
    conflicts: list[dict] = []
    unmatched: list[str] = []
    for row in rows_iter:
        if not row or row[0] is None:
            continue
        name = str(row[0])
        target = match_document(name, lookups)
        if target is None:
            unmatched.append(name)
            log.info(f"no manifest match for {name!r}, skipping")
            continue
        cc, case_id = target
        for group, model in GROUP_MODELS.items():
            values = {
                field: _clean(row[idx]) if idx < len(row) else ""
                for field, idx in group_cols[group].items()
                if field != "case_id"
            }
            seen = by_cc[model][cc].get(case_id)
            if seen is None:
                by_cc[model][cc][case_id] = values
                continue
            for field, alt in values.items():  # duplicate run of the same document
                if seen[field] != alt:
                    conflicts.append({
                        "model": model, "country": cc, "case_id": case_id,
                        "field": field, "kept": seen[field], "alternative": alt,
                    })
    if unmatched:
        log.warning(f"{len(unmatched)} document(s) had no manifest match: {unmatched[:5]} …")

    gold_indices = {
        cc: _gold_case_id_index(cc)
        for model_rows in by_cc.values()
        for cc in model_rows
    }
    for model, model_rows in by_cc.items():
        for cc, cases in sorted(model_rows.items()):
            index = gold_indices[cc]
            if index is None:
                log.info(f"[{cc}] no Goldenset on disk, skipping {len(cases)} Legora row(s)")
                continue
            out = inference_path(cc, prompt_version, source, model)
            out_rows: list[dict] = []
            dropped = 0
            for case_id, values in cases.items():
                gold = index.get(norm_case_id(case_id))
                if gold is None:
                    dropped += 1
                    log.info(f"[{cc}] manifest case_id {case_id!r} not in Goldenset, skipping")
                    continue
                out_row = {col: "" for col in columns}
                out_row["case_id"] = gold
                out_row["model"] = model
                out_row["inference_date"] = inference_date
                out_row.update(values)
                out_rows.append(out_row)
            write_inference_jsonl(out, out_rows, columns)
            log.info(f"[{cc}] wrote {len(out_rows)} {model} row(s) → {out} ({dropped} dropped)")

    conflicts.sort(key=lambda c: (c["model"], c["country"], c["case_id"], c["field"]))
    conflicts_out.parent.mkdir(parents=True, exist_ok=True)
    with open(conflicts_out, "w", encoding="utf-8") as f:
        for c in conflicts:
            f.write(json.dumps(c, ensure_ascii=False) + "\n")
    log.info(f"{len(conflicts)} duplicate-run conflict(s) -> {conflicts_out}")


def main() -> None:
    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s [%(levelname)s] %(message)s",
        handlers=[logging.StreamHandler(sys.stderr)],
    )
    parser = argparse.ArgumentParser(
        prog="legex-legora-ingest",
        description="Convert a Legora tabular-review export into per-country Goldenset_*_legora-{1,2}.jsonl files.",
    )
    parser.add_argument(
        "--xlsx",
        type=Path,
        default=settings.raw_dir / "legora_2026-08-01.xlsx",
        help="Path to the Legora export xlsx (default: data/raw/legora_2026-08-01.xlsx).",
    )
    parser.add_argument(
        "--manifest",
        type=Path,
        default=settings.raw_dir / "legora_bundle_manifest.csv",
        help="Bundle manifest mapping filenames to (country, case_id).",
    )
    parser.add_argument("--prompt_version", default="v3")
    parser.add_argument(
        "--source",
        choices=("full_text", "pdf"),
        default="full_text",
        help="Source bucket label used in the output filename (default: full_text).",
    )
    parser.add_argument(
        "--inference_date",
        default=None,
        help="ISO date the vendor ran the extraction (default: parsed from the xlsx filename).",
    )
    parser.add_argument("--conflicts", type=Path, default=CONFLICTS_PATH)
    args = parser.parse_args()
    ingest(
        xlsx=args.xlsx,
        manifest_csv=args.manifest,
        prompt_version=args.prompt_version,
        source=args.source,
        inference_date=args.inference_date,
        conflicts_out=args.conflicts,
    )


if __name__ == "__main__":
    main()