evaluation_kit/box_grouping/loading.py

"""Load annotation boxes and predicted rows from disk, and detect watermark rows."""

from __future__ import annotations

import argparse
import csv
import json
import sys
import unicodedata
from collections import Counter
from pathlib import Path
from typing import Any

from geometry import Box, rotated_rectangle_points, polygon_bounds
from models import (
    Word,
    AnnotationBox,
    PredictedRow,
    decomposable_container_box_ids,
    filter_redundant_annotation_boxes,
)


WATERMARK_PATTERNS = (
    {
        "tokens": ("national", "library", "of", "armenia", "ocr", "by", "portmind"),
        "token_counts": Counter(
            ("national", "library", "of", "armenia", "ocr", "by", "portmind")
        ),
        "distinctive_tokens": {"national", "library", "armenia", "ocr", "portmind"},
        "normalized_text": "nationallibraryofarmeniaocrbyportmind",
        "max_edit_distance": 2,
        "min_distinctive_token_matches": 3,
        "min_token_count": 6,
    },
    {
        "tokens": ("armenia", "ocr", "by", "portmind"),
        "token_counts": Counter(("armenia", "ocr", "by", "portmind")),
        "distinctive_tokens": {"armenia", "ocr", "portmind"},
        "normalized_text": "armeniaocrbyportmind",
        "max_edit_distance": 2,
        "min_distinctive_token_matches": 2,
        "min_token_count": 4,
    },
    {
        "tokens": ("national", "library", "of", "armenia"),
        "token_counts": Counter(("national", "library", "of", "armenia")),
        "distinctive_tokens": {"national", "library", "armenia"},
        "normalized_text": "nationallibraryofarmenia",
        "max_edit_distance": 2,
        "min_distinctive_token_matches": 2,
        "min_token_count": 3,
    },
)
WATERMARK_TOKEN_VOCABULARY = frozenset(
    token for pattern in WATERMARK_PATTERNS for token in pattern["tokens"]
)
NON_ARMENIAN_BOX_LETTER_RATIO_THRESHOLD = 0.9


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--annotations-json",
        type=Path,
        required=True,
        help="Path to the annotation JSON file.",
    )
    parser.add_argument(
        "--predictions-csv",
        type=Path,
        required=True,
        help="Path to the predicted word-box CSV file.",
    )
    parser.add_argument(
        "--coverage-threshold",
        type=float,
        default=1.0,
        help="Minimum fraction of words in a row that must fit a box for a full match.",
    )
    parser.add_argument(
        "--failure-example-count",
        type=int,
        default=5,
        help="Number of example failures to include per failure type in the report.",
    )
    parser.add_argument(
        "--filter",
        dest="filters",
        help=(
            "Optional comma-separated region filters, e.g. "
            "`non-armenian`, `graphics`, or `non-armenian,graphics`."
        ),
    )
    parser.add_argument(
        "--unit-level",
        dest="unit_level",
        choices=["word", "line"],
        default="word",
        help="Granularity of predicted rows: 'word' or 'line'.",
    )
    parser.add_argument(
        "--output",
        type=Path,
        help="Optional JSON path for the full report.",
    )
    return parser.parse_args()


def load_json(path: Path) -> Any:
    with path.open("r", encoding="utf-8") as handle:
        return json.load(handle)


def pick_geometry_result(results: list[dict[str, Any]]) -> dict[str, Any] | None:
    candidates = [
        result
        for result in results
        if {"x", "y", "width", "height"} <= set(result.get("value", {}).keys())
    ]
    if not candidates:
        return None

    def priority(result: dict[str, Any]) -> tuple[int, str]:
        from_name = result.get("from_name", "")
        if from_name == "label":
            return (0, from_name)
        return (1, from_name)

    return sorted(candidates, key=priority)[0]


def pick_transcription_result(results: list[dict[str, Any]]) -> dict[str, Any] | None:
    candidates = [
        result
        for result in results
        if result.get("from_name") == "transcription"
        and "text" in result.get("value", {})
    ]
    if not candidates:
        return None
    return candidates[0]


def text_from_result(result: dict[str, Any]) -> str:
    raw_text = result.get("value", {}).get("text", [])
    if isinstance(raw_text, list):
        return "\n".join(str(item) for item in raw_text)
    if raw_text is None:
        return ""
    return str(raw_text)


def parent_box_id_from_results(results: list[dict[str, Any]]) -> str | None:
    for result in results:
        if result.get("from_name") != "parent_id":
            continue
        raw_text = result.get("value", {}).get("text")
        if isinstance(raw_text, list) and raw_text:
            return str(raw_text[0])
    return None


def reading_order_from_results(results: list[dict[str, Any]]) -> int | None:
    for result in results:
        if result.get("from_name") != "reading_order":
            continue
        raw_text = result.get("value", {}).get("text")
        if isinstance(raw_text, list) and raw_text:
            try:
                return int(raw_text[0])
            except (TypeError, ValueError):
                return None
    return None


def labels_from_result(result: dict[str, Any]) -> tuple[str, ...]:
    raw_labels = result.get("value", {}).get("rectanglelabels", [])
    if not isinstance(raw_labels, list):
        return ()
    return tuple(str(label) for label in raw_labels if str(label).strip())


def letter_script(character: str) -> str | None:
    if not unicodedata.category(character).startswith("L"):
        return None
    character_name = unicodedata.name(character, "")
    if character_name.startswith("ARMENIAN"):
        return "armenian"
    if character_name.startswith("LATIN"):
        return "latin"
    if character_name.startswith("CYRILLIC"):
        return "cyrillic"
    return "other_letter"


def count_text_letter_scripts(text: str) -> Counter[str]:
    return Counter(
        script
        for character in text
        if (script := letter_script(character)) is not None
    )


def non_armenian_letter_ratio(text: str) -> tuple[int, int, float]:
    script_counts = count_text_letter_scripts(text)
    letter_count = sum(script_counts.values())
    latin_or_cyrillic_letter_count = (
        script_counts["latin"] + script_counts["cyrillic"]
    )
    ratio = (
        latin_or_cyrillic_letter_count / letter_count if letter_count else 0.0
    )
    return letter_count, latin_or_cyrillic_letter_count, ratio


def load_annotation_boxes(path: Path) -> list[AnnotationBox]:
    data = load_json(path)
    grouped_results: dict[str, list[dict[str, Any]]] = {}

    for annotation in data.get("annotations", []):
        for result in annotation.get("result", []):
            result_id = result.get("id")
            if result_id is None:
                continue
            grouped_results.setdefault(str(result_id), []).append(result)

    boxes: list[AnnotationBox] = []
    for box_id, results in grouped_results.items():
        transcription_result = pick_transcription_result(results)
        geometry_result = pick_geometry_result(results)
        if geometry_result is None:
            continue

        geometry = geometry_result["value"]
        x = float(geometry["x"])
        y = float(geometry["y"])
        width = float(geometry["width"])
        height = float(geometry["height"])
        rotation = float(geometry.get("rotation", 0.0))
        rect = Box(
            x_min=x,
            y_min=y,
            x_max=x + width,
            y_max=y + height,
        )
        polygon = rotated_rectangle_points(x, y, width, height, rotation)
        text = (
            text_from_result(transcription_result)
            if transcription_result is not None
            else ""
        )
        labels = labels_from_result(geometry_result)
        if "Rule" in labels:
            continue
        letter_count, latin_or_cyrillic_count, ratio = non_armenian_letter_ratio(
            text
        )
        boxes.append(
            AnnotationBox(
                box_id=box_id,
                rect=rect,
                text=text,
                has_transcription=transcription_result is not None,
                rotation=rotation,
                polygon=polygon,
                bounds=polygon_bounds(polygon),
                labels=labels,
                letter_count=letter_count,
                latin_or_cyrillic_letter_count=latin_or_cyrillic_count,
                non_armenian_letter_ratio=ratio,
                excluded_as_non_armenian_text=(
                    ratio > NON_ARMENIAN_BOX_LETTER_RATIO_THRESHOLD
                ),
                parent_box_id=parent_box_id_from_results(results),
                reading_order=reading_order_from_results(results),
            )
        )

    decomposable_ids = decomposable_container_box_ids(boxes)
    boxes = [box for box in boxes if box.box_id not in decomposable_ids]

    return sorted(
        filter_redundant_annotation_boxes(boxes),
        key=lambda item: (item.bounds.y_min, item.bounds.x_min, item.box_id),
    )


def load_predicted_rows(path: Path, unit_level: str = "word") -> list[PredictedRow]:
    """Load predicted word boxes from an evaluation CSV, grouped into rows.

    For ``unit_level="line"``, rows sharing the same ``group_row`` value are
    assembled into one PredictedRow (a full text line built from its
    constituent words) — the CSV's own row grouping is authoritative.

    For ``unit_level="word"`` (the default), ``group_row`` is ignored and
    every CSV row becomes its own single-word PredictedRow. This matters for
    CSVs where multiple individual word boxes share a ``group_row`` because
    an upstream step assigned them to the same visual line (e.g. raw
    box_grouping/data CSVs): without this, a word-level evaluation would
    incorrectly treat that whole line as one atomic unit instead of scoring
    each word independently against the ground truth.
    """
    grouped_words: dict[str, list[Word]] = {}

    with path.open("r", encoding="utf-8", newline="") as handle:
        reader = csv.DictReader(handle)
        for index, row in enumerate(reader):
            row_id = str(index) if unit_level == "word" else row["group_row"]
            word = Word(
                box=Box(
                    x_min=float(row["x1"]),
                    y_min=float(row["y1"]),
                    x_max=float(row["x2"]),
                    y_max=float(row["y2"]),
                ),
                text=row.get("text", ""),
            )
            grouped_words.setdefault(row_id, []).append(word)

    return [
        PredictedRow(
            row_id=row_id,
            words=sorted(words, key=lambda w: (w.box.x_min, w.box.y_min)),
        )
        for row_id, words in grouped_words.items()
    ]


def normalize_token(text: str) -> str:
    return "".join(character.lower() for character in text if character.isalnum())


def normalized_token_segments(text: str) -> list[str]:
    segments: list[str] = []
    current_characters: list[str] = []
    for character in text:
        if character.isalnum():
            current_characters.append(character)
            continue
        if current_characters:
            normalized = normalize_token("".join(current_characters))
            if normalized:
                segments.append(normalized)
            current_characters = []
    if current_characters:
        normalized = normalize_token("".join(current_characters))
        if normalized:
            segments.append(normalized)
    return segments


def matches_watermark_pattern(
    normalized_tokens: list[str], pattern: dict[str, Any]
) -> bool:
    # Deferred import: text_metrics lives in evaluation/, not box_grouping/.
    # Avoids a top-level cycle (loading → text_metrics → spatial → loading).
    _evaluation = str(Path(__file__).resolve().parent.parent / "evaluation")
    if _evaluation not in sys.path:
        sys.path.insert(0, _evaluation)
    from text_metrics import edit_distance

    token_counts = Counter(normalized_tokens)
    distinctive_token_matches = len(
        set(token_counts) & pattern["distinctive_tokens"]
    )
    if distinctive_token_matches < pattern["min_distinctive_token_matches"]:
        return False

    if len(normalized_tokens) >= pattern["min_token_count"] and all(
        count <= pattern["token_counts"].get(token, 0)
        for token, count in token_counts.items()
    ):
        return True

    normalized_text = "".join(normalized_tokens)
    if (
        abs(len(normalized_text) - len(pattern["normalized_text"]))
        > pattern["max_edit_distance"]
    ):
        return False

    return (
        edit_distance(normalized_text, pattern["normalized_text"])
        <= pattern["max_edit_distance"]
    )


def normalized_row_tokens(predicted_row: PredictedRow) -> list[str]:
    normalized_tokens: list[str] = []
    for word in predicted_row.words:
        normalized_tokens.extend(normalized_token_segments(word.text))
    return normalized_tokens


def is_watermark_row(predicted_row: PredictedRow) -> bool:
    normalized_tokens = normalized_row_tokens(predicted_row)
    if not normalized_tokens:
        return False
    return any(
        matches_watermark_pattern(normalized_tokens, pattern)
        for pattern in WATERMARK_PATTERNS
    )


def matches_watermark_subsequence(normalized_tokens: list[str]) -> bool:
    if not normalized_tokens:
        return False
    token_count = len(normalized_tokens)
    for pattern in WATERMARK_PATTERNS:
        pattern_tokens = pattern["tokens"]
        if token_count > len(pattern_tokens):
            continue
        for start_index in range(len(pattern_tokens) - token_count + 1):
            if (
                tuple(normalized_tokens)
                == pattern_tokens[start_index : start_index + token_count]
            ):
                return True
    return False


def watermark_row_ids(
    predicted_rows: list[PredictedRow],
    local_boxes_share_line_fn: Any,
) -> set[str]:
    """Return the set of row_ids that are part of a watermark (direct or fragmented)."""
    row_details = [
        {
            "predicted_row": predicted_row,
            "row_box": _row_box(predicted_row),
            "normalized_tokens": normalized_row_tokens(predicted_row),
            "is_direct_watermark": is_watermark_row(predicted_row),
        }
        for predicted_row in predicted_rows
    ]
    ignored_row_ids = {
        detail["predicted_row"].row_id
        for detail in row_details
        if detail["is_direct_watermark"]
    }
    candidate_indexes = [
        index
        for index, detail in enumerate(row_details)
        if detail["is_direct_watermark"]
        or (
            detail["normalized_tokens"]
            and all(
                token in WATERMARK_TOKEN_VOCABULARY
                for token in detail["normalized_tokens"]
            )
        )
    ]

    visited_indexes: set[int] = set()
    for candidate_index in candidate_indexes:
        if candidate_index in visited_indexes:
            continue

        component_indexes: list[int] = []
        pending_indexes = [candidate_index]
        visited_indexes.add(candidate_index)
        while pending_indexes:
            current_index = pending_indexes.pop()
            component_indexes.append(current_index)
            current_box = row_details[current_index]["row_box"]
            for other_index in candidate_indexes:
                if other_index in visited_indexes:
                    continue
                if not local_boxes_share_line_fn(
                    current_box, row_details[other_index]["row_box"]
                ):
                    continue
                visited_indexes.add(other_index)
                pending_indexes.append(other_index)

        if len(component_indexes) < 2:
            continue

        ordered_component_indexes = sorted(
            component_indexes,
            key=lambda i: (
                row_details[i]["row_box"].x_min,
                row_details[i]["row_box"].y_min,
                row_details[i]["predicted_row"].row_id,
            ),
        )
        combined_tokens: list[str] = []
        for component_index in ordered_component_indexes:
            combined_tokens.extend(row_details[component_index]["normalized_tokens"])

        if any(
            matches_watermark_pattern(combined_tokens, pattern)
            for pattern in WATERMARK_PATTERNS
        ) or matches_watermark_subsequence(combined_tokens):
            ignored_row_ids.update(
                row_details[component_index]["predicted_row"].row_id
                for component_index in component_indexes
            )

    return ignored_row_ids


def _row_box(predicted_row: PredictedRow) -> Box:
    """Compute the bounding box of a predicted row without importing from spatial."""
    return Box(
        x_min=min(word.box.x_min for word in predicted_row.words),
        y_min=min(word.box.y_min for word in predicted_row.words),
        x_max=max(word.box.x_max for word in predicted_row.words),
        y_max=max(word.box.y_max for word in predicted_row.words),
    )