mindnet/app/core/chunking/chunking_strategies.py

"""
FILE: app/core/chunking/chunking_strategies.py
DESCRIPTION: Strategien für atomares Sektions-Chunking v3.4.1.
             Garantiert Sektions-Integrität (Atomic Units) durch Look-Ahead.
"""
import math
from typing import List, Dict, Any, Optional
from .chunking_models import RawBlock, Chunk
from .chunking_utils import estimate_tokens
from .chunking_parser import split_sentences

def _safe_estimate_tokens(text: str) -> int:
    """Konservative Schätzung für MD und deutsche Texte (len/2.8)."""
    return max(1, math.ceil(len(text.strip()) / 2.8))

def _create_context_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
    """Baut den Breadcrumb-Kontext für das Embedding-Fenster (H1 > H2)."""
    parts = []
    if doc_title: parts.append(doc_title)
    if sec_title and sec_title != doc_title: parts.append(sec_title)
    prefix = " > ".join(parts)
    return f"{prefix}\n{text}".strip() if prefix else text

def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, doc_title: str = "") -> List[Chunk]:
    """
    Sektions-Chunking: Behandelt Abschnitte als unteilbare Einheiten.
    Schiebt ganze Abschnitte in den nächsten Chunk, falls das Limit erreicht ist.
    """
    strict = config.get("strict_heading_split", False)
    target = config.get("target", 400)
    max_tokens = config.get("max", 600)
    split_level = config.get("split_level", 2)
    overlap_cfg = config.get("overlap", (50, 80))
    overlap = sum(overlap_cfg) // 2 if isinstance(overlap_cfg, (list, tuple)) else overlap_cfg
    
    chunks: List[Chunk] = []

    def _emit_chunk(block_list: List[RawBlock]):
        """Erzeugt ein finales Chunk-Objekt aus einer Liste von Blöcken."""
        if not block_list: return
        txt = "\n\n".join([b.text for b in block_list])
        idx = len(chunks)
        # Metadaten vom ersten Block der Gruppe (Header)
        title = block_list[0].section_title
        path = block_list[0].section_path
        win = _create_context_win(doc_title, title, txt)
        chunks.append(Chunk(
            id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx,
            text=txt, window=win, token_count=_safe_estimate_tokens(txt),
            section_title=title, section_path=path,
            neighbors_prev=None, neighbors_next=None
        ))

    def _emit_split_section(sec_blocks: List[RawBlock]):
        """Splittet eine einzelne Sektion, die für sich allein zu groß ist."""
        full_text = "\n\n".join([b.text for b in sec_blocks])
        main_title = sec_blocks[0].section_title
        main_path = sec_blocks[0].section_path
        header_text = sec_blocks[0].text if sec_blocks[0].kind == "heading" else ""
        
        sents = split_sentences(full_text)
        cur_sents = []; sub_len = 0
        
        for s in sents:
            slen = _safe_estimate_tokens(s)
            if sub_len + slen > target and cur_sents:
                _emit_chunk([RawBlock("paragraph", " ".join(cur_sents), None, main_path, main_title)])
                # Header Injection für den Kontext im nächsten Teil-Chunk
                ov_s = [header_text] if header_text else []
                ov_l = _safe_estimate_tokens(header_text) if header_text else 0
                for os in reversed(cur_sents):
                    if os == header_text: continue
                    t_len = _safe_estimate_tokens(os)
                    if ov_l + t_len < overlap:
                        ov_s.insert(len(ov_s)-1 if header_text else 0, os)
                        ov_l += t_len
                    else: break
                cur_sents = list(ov_s); cur_sents.append(s); sub_len = ov_l + slen
            else:
                cur_sents.append(s); sub_len += slen
        if cur_sents:
            _emit_chunk([RawBlock("paragraph", " ".join(cur_sents), None, main_path, main_title)])

    # SCHRITT 1: Gruppierung in atomare Einheiten (Sektionen)
    sections: List[List[RawBlock]] = []
    curr_sec: List[RawBlock] = []
    for b in blocks:
        if b.kind == "heading" and b.level <= split_level:
            if curr_sec: sections.append(curr_sec)
            curr_sec = [b]
        else:
            curr_sec.append(b)
    if curr_sec: sections.append(curr_sec)

    # SCHRITT 2: Verarbeitung der Sektionen mit strengem Look-Ahead
    current_chunk_buf = []
    current_tokens = 0

    for sec in sections:
        sec_text = "\n\n".join([b.text for b in sec])
        sec_tokens = _safe_estimate_tokens(sec_text)
        
        if current_chunk_buf:
            # PRÜFUNG: Würde die neue Sektion den aktuellen Chunk sprengen?
            # ODER: Haben wir das Target bereits erreicht und fangen lieber neu an?
            if (current_tokens + sec_tokens > max_tokens) or (current_tokens >= target):
                _emit_chunk(current_chunk_buf)
                current_chunk_buf = []
                current_tokens = 0
            # PRÜFUNG: Harter Split gefordert an Überschriften
            elif strict and sec[0].kind == "heading" and sec[0].level == split_level:
                _emit_chunk(current_chunk_buf)
                current_chunk_buf = []
                current_tokens = 0

        # Wenn eine EINZELNE Sektion alleine schon das Limit sprengt
        if sec_tokens > max_tokens:
            if current_chunk_buf:
                _emit_chunk(current_chunk_buf)
                current_chunk_buf = []
                current_tokens = 0
            _emit_split_section(sec)
        else:
            current_chunk_buf.extend(sec)
            current_tokens += sec_tokens + 2 # +2 für Newline Join

    # Letzten Puffer schreiben
    if current_chunk_buf:
        _emit_chunk(current_chunk_buf)

    return chunks

def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, context_prefix: str = "") -> List[Chunk]:
    """Basis-Sliding-Window für flache Texte ohne Sektionsfokus."""
    target = config.get("target", 400)
    max_tokens = config.get("max", 600)
    chunks: List[Chunk] = []
    buf: List[RawBlock] = []
    
    for b in blocks:
        b_tokens = _safe_estimate_tokens(b.text)
        current_tokens = sum(_safe_estimate_tokens(x.text) for x in buf) if buf else 0
        if current_tokens + b_tokens > max_tokens and buf:
            txt = "\n\n".join([x.text for x in buf])
            idx = len(chunks)
            win = f"{context_prefix}\n{txt}".strip() if context_prefix else txt
            chunks.append(Chunk(id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx, text=txt, window=win, token_count=current_tokens, section_title=buf[0].section_title, section_path=buf[0].section_path, neighbors_prev=None, neighbors_next=None))
            buf = []
        buf.append(b)
    if buf:
        txt = "\n\n".join([x.text for x in buf]); idx = len(chunks)
        win = f"{context_prefix}\n{txt}".strip() if context_prefix else txt
        chunks.append(Chunk(id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx, text=txt, window=win, token_count=_safe_estimate_tokens(txt), section_title=buf[0].section_title, section_path=buf[0].section_path, neighbors_prev=None, neighbors_next=None))
    return chunks