""" FILE: app/core/chunking/chunking_strategies.py DESCRIPTION: Strategien für atomares Sektions-Chunking v3.9.9. Implementiert das 'Pack-and-Carry-Over' Verfahren nach Regel 1-3. - Keine redundante Kanten-Injektion. - Strikte Einhaltung von Sektionsgrenzen via Look-Ahead. - Fix: Synchronisierung der Parameter mit dem Orchestrator (context_prefix). WP-24c v4.2.5: Strict-Mode ohne Carry-Over - Bei strict_heading_split wird nach jeder Sektion geflasht. WP-26 v1.0: section_type und block_id werden an Chunks weitergegeben. """ 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 _create_win(context_prefix: str, sec_title: Optional[str], text: str) -> str: """Baut den Breadcrumb-Kontext für das Embedding-Fenster.""" parts = [context_prefix] if context_prefix else [] # Verhindert Dopplung, falls der Context-Prefix (H1) bereits den Sektionsnamen enthält if sec_title and f"# {sec_title}" != context_prefix and sec_title not in (context_prefix or ""): 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, context_prefix: str = "") -> List[Chunk]: """ Universelle Heading-Strategie mit Carry-Over Logik. Synchronisiert auf context_prefix für Kompatibilität mit dem Orchestrator. """ smart_edge = config.get("enable_smart_edge_allocation", True) 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(txt, title, path, section_type=None, block_id=None): """ Schreibt den finalen Chunk ohne Text-Modifikationen. WP-26 v1.0: Erweitert um section_type und block_id. """ idx = len(chunks) win = _create_win(context_prefix, title, txt) chunks.append(Chunk( id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx, text=txt, window=win, token_count=estimate_tokens(txt), section_title=title, section_path=path, neighbors_prev=None, neighbors_next=None, section_type=section_type, block_id=block_id )) # --- SCHRITT 1: Gruppierung in atomare Sektions-Einheiten --- # WP-26 v1.0: Erweitert um section_type und block_id Tracking sections: List[Dict[str, Any]] = [] curr_blocks = [] for b in blocks: if b.kind == "heading" and b.level <= split_level: if curr_blocks: # WP-26 v1.0: Finde den effektiven section_type und block_id für diese Sektion # Priorisiere den ersten Block mit section_type, sonst den Heading-Block effective_section_type = None effective_block_id = None for cb in curr_blocks: if cb.section_type and effective_section_type is None: effective_section_type = cb.section_type if cb.block_id and effective_block_id is None: effective_block_id = cb.block_id sections.append({ "text": "\n\n".join([x.text for x in curr_blocks]), "meta": curr_blocks[0], "is_empty": len(curr_blocks) == 1 and curr_blocks[0].kind == "heading", "section_type": effective_section_type, "block_id": effective_block_id }) curr_blocks = [b] else: curr_blocks.append(b) if curr_blocks: # WP-26 v1.0: Gleiche Logik für den letzten Block effective_section_type = None effective_block_id = None for cb in curr_blocks: if cb.section_type and effective_section_type is None: effective_section_type = cb.section_type if cb.block_id and effective_block_id is None: effective_block_id = cb.block_id sections.append({ "text": "\n\n".join([x.text for x in curr_blocks]), "meta": curr_blocks[0], "is_empty": len(curr_blocks) == 1 and curr_blocks[0].kind == "heading", "section_type": effective_section_type, "block_id": effective_block_id }) # --- SCHRITT 2: Verarbeitung der Queue --- queue = list(sections) current_chunk_text = "" # WP-26 v1.0: Erweitert um section_type und block_id current_meta = {"title": None, "path": "/", "section_type": None, "block_id": None} # Bestimmung des Modus: Hard-Split wenn smart_edge=False ODER strict=True is_hard_split_mode = (not smart_edge) or (strict) while queue: item = queue.pop(0) item_text = item["text"] # Initialisierung für neuen Chunk if not current_chunk_text: current_meta["title"] = item["meta"].section_title current_meta["path"] = item["meta"].section_path # WP-26 v1.0: section_type und block_id aus Item übernehmen current_meta["section_type"] = item.get("section_type") current_meta["block_id"] = item.get("block_id") # FALL A: HARD SPLIT MODUS (WP-24c v4.2.5: Strict-Mode ohne Carry-Over) if is_hard_split_mode: # WP-24c v4.2.5: Bei strict_heading_split: true wird nach JEDER Sektion geflasht # Kein Carry-Over erlaubt, auch nicht für leere Überschriften if current_chunk_text: # Flashe vorherigen Chunk _emit(current_chunk_text, current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) current_chunk_text = "" # Neue Sektion: Initialisiere Meta current_meta["title"] = item["meta"].section_title current_meta["path"] = item["meta"].section_path # WP-26 v1.0: section_type und block_id aus Item übernehmen current_meta["section_type"] = item.get("section_type") current_meta["block_id"] = item.get("block_id") # WP-24c v4.2.5: Auch leere Sektionen werden als separater Chunk erstellt # (nur Überschrift, kein Inhalt) if item.get("is_empty", False): # Leere Sektion: Nur Überschrift als Chunk _emit(item_text, current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) else: # Normale Sektion: Prüfe auf Token-Limit if estimate_tokens(item_text) > max_tokens: # Sektion zu groß: Smart Zerlegung (aber trotzdem in separaten Chunks) sents = split_sentences(item_text) header_prefix = item["meta"].text if item["meta"].kind == "heading" else "" take_sents = []; take_len = 0 while sents: s = sents.pop(0); slen = estimate_tokens(s) if take_len + slen > target and take_sents: _emit(" ".join(take_sents), current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) take_sents = [s]; take_len = slen else: take_sents.append(s); take_len += slen if take_sents: _emit(" ".join(take_sents), current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) else: # Sektion passt: Direkt als Chunk _emit(item_text, current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) current_chunk_text = "" continue # FALL B: SMART MODE (Regel 1-3) combined_text = (current_chunk_text + "\n\n" + item_text).strip() if current_chunk_text else item_text combined_est = estimate_tokens(combined_text) if combined_est <= max_tokens: # Regel 1 & 2: Passt rein laut Schätzung -> Aufnehmen current_chunk_text = combined_text else: if current_chunk_text: # Regel 2: Flashen an Sektionsgrenze, Item zurücklegen _emit(current_chunk_text, current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) current_chunk_text = "" queue.insert(0, item) else: # Regel 3: Einzelne Sektion zu groß -> Smart Zerlegung sents = split_sentences(item_text) header_prefix = item["meta"].text if item["meta"].kind == "heading" else "" take_sents = []; take_len = 0 while sents: s = sents.pop(0); slen = estimate_tokens(s) if take_len + slen > target and take_sents: sents.insert(0, s); break take_sents.append(s); take_len += slen _emit(" ".join(take_sents), current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) if sents: remainder = " ".join(sents) # Kontext-Erhalt: Überschrift für den Rest wiederholen if header_prefix and not remainder.startswith(header_prefix): remainder = header_prefix + "\n\n" + remainder # Carry-Over: Rest wird vorne in die Queue geschoben # WP-26 v1.0: section_type und block_id weitergeben queue.insert(0, {"text": remainder, "meta": item["meta"], "is_split": True, "section_type": item.get("section_type"), "block_id": item.get("block_id")}) if current_chunk_text: _emit(current_chunk_text, current_meta["title"], current_meta["path"], current_meta["section_type"], current_meta["block_id"]) return chunks def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, context_prefix: str = "") -> List[Chunk]: """ Standard-Sliding-Window für flache Texte ohne Sektionsfokus. WP-26 v1.0: Erweitert um section_type und block_id Weitergabe. """ target = config.get("target", 400); max_tokens = config.get("max", 600) chunks: List[Chunk] = []; buf: List[RawBlock] = [] for b in blocks: b_tokens = estimate_tokens(b.text) curr_tokens = sum(estimate_tokens(x.text) for x in buf) if buf else 0 if curr_tokens + b_tokens > max_tokens and buf: txt = "\n\n".join([x.text for x in buf]); idx = len(chunks) win = _create_win(context_prefix, buf[0].section_title, txt) # WP-26 v1.0: Finde effektiven section_type und block_id effective_section_type = next((b.section_type for b in buf if b.section_type), None) effective_block_id = next((b.block_id for b in buf if b.block_id), None) chunks.append(Chunk( id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx, text=txt, window=win, token_count=curr_tokens, section_title=buf[0].section_title, section_path=buf[0].section_path, neighbors_prev=None, neighbors_next=None, section_type=effective_section_type, block_id=effective_block_id )) buf = [] buf.append(b) if buf: txt = "\n\n".join([x.text for x in buf]); idx = len(chunks) win = _create_win(context_prefix, buf[0].section_title, txt) # WP-26 v1.0: Finde effektiven section_type und block_id effective_section_type = next((b.section_type for b in buf if b.section_type), None) effective_block_id = next((b.block_id for b in buf if b.block_id), None) chunks.append(Chunk( id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx, text=txt, window=win, token_count=estimate_tokens(txt), section_title=buf[0].section_title, section_path=buf[0].section_path, neighbors_prev=None, neighbors_next=None, section_type=effective_section_type, block_id=effective_block_id )) return chunks