"""Sentence-aware document segmenter.

Splits documents into chunks that respect sentence boundaries, preserve
source character offsets, and use document-type-specific strategies.

Key invariant: chunk.text == source_text[chunk.start_char:chunk.end_char]
"""

from __future__ import annotations

import re

from services.intelligence_pipeline_v3.segmenter.boilerplate import score_boilerplate
from services.intelligence_pipeline_v3.segmenter.models import DocumentChunk
from services.intelligence_pipeline_v3.segmenter.strategies import (
    ChunkStrategy,
    get_strategy,
)

# Sentence-ending patterns: period/question/exclamation followed by space or newline
_SENTENCE_END = re.compile(r"(?<=[.!?])\s+")

# Speaker turn pattern for transcripts
_SPEAKER_PATTERN = re.compile(
    r"^([A-Z][a-zA-Z\s\-\.]+(?:\s*[-–—]\s*[A-Za-z\s,]+)?)\s*:\s*",
    re.MULTILINE,
)

# Filing section headers
_FILING_SECTION_PATTERN = re.compile(
    r"^((?:Item\s+\d+[A-Z]?[\.\:]|PART\s+[IVX]+)[^\n]*)",
    re.IGNORECASE | re.MULTILINE,
)


class Segmenter:
    """Sentence-aware document segmenter with document-type-specific strategies.

    Does NOT truncate documents — processes full text regardless of length.
    Each chunk preserves exact character offsets into the source document.
    """

    def segment(
        self,
        text: str,
        document_type: str,
        document_id: str = "",
    ) -> list[DocumentChunk]:
        """Segment a document into chunks.

        Args:
            text: Full document text (no truncation applied).
            document_type: Type of document for strategy selection.
            document_id: Identifier for deterministic chunk IDs.

        Returns:
            List of DocumentChunk with valid offsets and checksums.
        """
        if not text:
            return []

        strategy = get_strategy(document_type)

        # Get structural boundaries based on document type
        boundaries = self._find_boundaries(text, strategy, document_type)

        # Build chunks respecting sentence boundaries and strategy limits
        chunks = self._build_chunks(
            text=text,
            boundaries=boundaries,
            strategy=strategy,
            document_type=document_type,
            document_id=document_id,
        )

        # Apply overlap between adjacent chunks
        chunks = self._apply_overlap(chunks, text, strategy, document_id, document_type)

        # Score boilerplate
        for chunk in chunks:
            chunk.boilerplate_score = score_boilerplate(chunk.text)

        # Apply section path and speaker metadata
        self._apply_metadata(chunks, text, document_type)

        return chunks

    def _find_boundaries(
        self,
        text: str,
        strategy: ChunkStrategy,
        document_type: str,
    ) -> list[int]:
        """Find structural boundary positions in the text.

        Returns sorted list of character offsets where structural breaks occur.
        """
        boundaries: set[int] = set()

        for pattern in strategy.boundary_patterns:
            for match in pattern.finditer(text):
                boundaries.add(match.start())

        return sorted(boundaries)

    def _find_sentence_boundaries(self, text: str, start: int, end: int) -> list[int]:
        """Find sentence-ending positions within a text range.

        Returns character offsets (relative to full document) after sentence-ending punctuation.
        """
        segment = text[start:end]
        positions: list[int] = []
        for match in _SENTENCE_END.finditer(segment):
            positions.append(start + match.start())
        return positions

    def _build_chunks(
        self,
        text: str,
        boundaries: list[int],
        strategy: ChunkStrategy,
        document_type: str,
        document_id: str,
    ) -> list[DocumentChunk]:
        """Build initial chunks from text using boundaries and sentence awareness."""
        chunks: list[DocumentChunk] = []
        text_len = len(text)
        pos = 0

        while pos < text_len:
            # Determine the ideal end position
            ideal_end = min(pos + strategy.target_chars, text_len)
            max_end = min(pos + strategy.max_chars, text_len)

            if ideal_end >= text_len:
                # Last chunk — take everything remaining
                chunk_end = text_len
            else:
                # Try to break at a structural boundary between ideal and max
                chunk_end = self._find_best_break(
                    text, pos, ideal_end, max_end, boundaries, strategy
                )

            chunk_text = text[pos:chunk_end]

            # Skip empty chunks (shouldn't happen, but defensive)
            if not chunk_text.strip():
                pos = chunk_end
                continue

            chunk = DocumentChunk(
                chunk_id=f"{document_id}:{pos}",
                document_id=document_id,
                document_type=document_type,
                section_path=[],
                speaker=None,
                start_char=pos,
                end_char=chunk_end,
                text=chunk_text,
                overlap_left=0,
                overlap_right=0,
                boilerplate_score=0.0,
            )
            chunks.append(chunk)
            pos = chunk_end

        return chunks

    def _find_best_break(
        self,
        text: str,
        start: int,
        ideal_end: int,
        max_end: int,
        boundaries: list[int],
        strategy: ChunkStrategy,
    ) -> int:
        """Find the best break point between ideal_end and max_end.

        Priority:
        1. Structural boundary nearest to ideal_end (within target..max range)
        2. Sentence boundary nearest to ideal_end
        3. Whitespace nearest to ideal_end
        4. Hard cut at ideal_end
        """
        # Look for structural boundaries in the window [ideal_end - target_chars/4, max_end]
        search_start = max(start, ideal_end - strategy.target_chars // 4)
        best_structural = None
        for b in boundaries:
            if search_start <= b <= max_end and b > start:
                if best_structural is None or abs(b - ideal_end) < abs(best_structural - ideal_end):
                    best_structural = b
        if best_structural is not None:
            return best_structural

        # Look for sentence boundaries near ideal_end
        sentence_breaks = self._find_sentence_boundaries(text, search_start, max_end)
        if sentence_breaks:
            # Pick the one closest to ideal_end
            best_sentence = min(sentence_breaks, key=lambda s: abs(s - ideal_end))
            # Use position after the sentence-ending whitespace
            after = best_sentence
            while after < max_end and text[after] in " \t\n\r":
                after += 1
            return after

        # Fall back to whitespace break
        search_region = text[ideal_end:max_end]
        ws_match = re.search(r"\s+", search_region)
        if ws_match:
            return ideal_end + ws_match.end()

        # Hard cut
        return ideal_end

    def _apply_overlap(
        self,
        chunks: list[DocumentChunk],
        text: str,
        strategy: ChunkStrategy,
        document_id: str,
        document_type: str,
    ) -> list[DocumentChunk]:
        """Apply overlap between adjacent chunks by extending start/end.

        Overlap is achieved by moving each chunk's start_char backward
        to include trailing content from the previous chunk.
        """
        if len(chunks) <= 1 or strategy.overlap_chars == 0:
            return chunks

        result: list[DocumentChunk] = []
        for i, chunk in enumerate(chunks):
            new_start = chunk.start_char
            new_end = chunk.end_char
            overlap_left = 0
            overlap_right = 0

            if i > 0:
                # Extend start backward for overlap
                overlap_target = min(strategy.overlap_chars, chunk.start_char)
                new_start = max(0, chunk.start_char - overlap_target)

                # Snap to sentence boundary if possible
                if new_start < chunk.start_char:
                    region = text[new_start:chunk.start_char]
                    sentence_matches = list(_SENTENCE_END.finditer(region))
                    if sentence_matches:
                        # Use the latest sentence break in the overlap region
                        last_match = sentence_matches[-1]
                        candidate = new_start + last_match.end()
                        if candidate < chunk.start_char:
                            new_start = candidate

                overlap_left = chunk.start_char - new_start

            if i < len(chunks) - 1:
                # Calculate how much the next chunk will overlap into this one
                next_chunk = chunks[i + 1]
                overlap_target = min(strategy.overlap_chars, len(text) - next_chunk.start_char)
                potential_overlap_start = max(0, next_chunk.start_char - overlap_target)

                # The overlap_right for this chunk = how much the next chunk's
                # overlap will include from this chunk's content
                overlap_right = chunk.end_char - max(potential_overlap_start, chunk.start_char)
                overlap_right = max(0, overlap_right)

            new_text = text[new_start:new_end]
            if not new_text.strip():
                result.append(chunk)
                continue

            result.append(DocumentChunk(
                chunk_id=f"{document_id}:{new_start}",
                document_id=document_id,
                document_type=document_type,
                section_path=chunk.section_path,
                speaker=chunk.speaker,
                start_char=new_start,
                end_char=new_end,
                text=new_text,
                overlap_left=overlap_left,
                overlap_right=overlap_right,
                boilerplate_score=chunk.boilerplate_score,
            ))

        return result

    def _apply_metadata(
        self,
        chunks: list[DocumentChunk],
        text: str,
        document_type: str,
    ) -> None:
        """Apply section_path and speaker metadata to chunks in place."""
        if document_type.lower() == "transcript":
            self._apply_speaker_metadata(chunks, text)
        elif document_type.lower() == "filing":
            self._apply_filing_sections(chunks, text)

    def _apply_speaker_metadata(
        self,
        chunks: list[DocumentChunk],
        text: str,
    ) -> None:
        """Find speaker turns and assign speaker labels to transcript chunks."""
        speakers: list[tuple[int, str]] = []
        for match in _SPEAKER_PATTERN.finditer(text):
            speakers.append((match.start(), match.group(1).strip()))

        if not speakers:
            return

        for chunk in chunks:
            # Find the most recent speaker before or within this chunk
            current_speaker = None
            for pos, name in speakers:
                if pos <= chunk.end_char:
                    if pos < chunk.start_char:
                        current_speaker = name
                    else:
                        current_speaker = name
                else:
                    break
            chunk.speaker = current_speaker

    def _apply_filing_sections(
        self,
        chunks: list[DocumentChunk],
        text: str,
    ) -> None:
        """Find filing section headers and assign section_path to chunks."""
        sections: list[tuple[int, str]] = []
        for match in _FILING_SECTION_PATTERN.finditer(text):
            sections.append((match.start(), match.group(1).strip()))

        if not sections:
            return

        for chunk in chunks:
            # Build section path from all sections that precede or start within this chunk
            path: list[str] = []
            for pos, title in sections:
                if pos < chunk.end_char:
                    # Keep track of the most recent section(s)
                    if pos <= chunk.start_char:
                        path = [title]
                    else:
                        path.append(title)
                else:
                    break
            chunk.section_path = path