mindnet/app/core/chunking/chunking_strategies.py
Lars 52fdc425f7 Enhance chunking strategies and graph utilities for section-type transitions and block ID extraction
- Implemented WP-26 v1.1: Section-Type-Wechsel erzwingt Split auch in SMART MODE (Schritt 2) zur Verbesserung der Chunking-Logik.
- Updated `parse_link_target` to extract block IDs from section strings, ensuring accurate handling of links with block references.
- Added unit tests to validate section-type change behavior and block ID extraction functionality, enhancing overall reliability.
2026-01-25 17:47:22 +01:00

322 lines
16 KiB
Python

"""
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.
WP-26 v1.1: Section-Type-Wechsel erzwingt IMMER einen neuen Chunk (unabhängig vom Profil).
Gleiche section_types folgen dem normalen Chunking-Verhalten.
"""
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
# WP-26 v1.1: Section-Type-Wechsel erzwingt IMMER eine neue Sektion (unabhängig vom Profil)
sections: List[Dict[str, Any]] = []
curr_blocks = []
current_section_type = None # WP-26 v1.1: Tracking des aktuellen section_type
def _flush_section():
"""Hilfsfunktion zum Abschließen einer Sektion."""
nonlocal curr_blocks, current_section_type
if not curr_blocks:
return
# 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 = []
current_section_type = effective_section_type
for b in blocks:
# WP-26 v1.1: Prüfe auf Section-Type-Wechsel BEVOR wir den Block hinzufügen
# Wenn sich der section_type ändert, muss die aktuelle Sektion abgeschlossen werden
block_section_type = b.section_type
# Section-Type-Wechsel erkannt?
# (Wechsel ist: alter Typ != neuer Typ UND mindestens einer ist nicht None)
is_section_type_change = (
curr_blocks and # Es gibt bereits Blöcke
block_section_type is not None and # Neuer Block hat expliziten section_type
current_section_type != block_section_type # Typ hat sich geändert
)
if b.kind == "heading" and b.level <= split_level:
# Heading-basierter Split (Standard-Verhalten)
_flush_section()
curr_blocks = [b]
current_section_type = block_section_type # Update tracking
elif is_section_type_change:
# WP-26 v1.1: Section-Type-Wechsel erzwingt neuen Chunk
_flush_section()
curr_blocks = [b]
current_section_type = block_section_type
else:
# Normales Hinzufügen zum aktuellen Block
curr_blocks.append(b)
# Update section_type wenn Block einen hat und wir noch keinen haben
if block_section_type and not current_section_type:
current_section_type = block_section_type
# Letzte Sektion abschließen
_flush_section()
# --- 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)
# WP-26 v1.1: Prüfe auf Section-Type-Wechsel AUCH in Schritt 2
# Wenn sich der section_type zwischen current_meta und item ändert, muss gesplittet werden
item_section_type = item.get("section_type")
current_section_type_meta = current_meta.get("section_type")
# Section-Type-Wechsel: Von None zu einem Typ ODER von einem Typ zu einem anderen
is_section_type_change_step2 = (
current_chunk_text and # Es gibt bereits Content
(
# Wechsel von None zu einem Typ
(current_section_type_meta is None and item_section_type is not None) or
# Wechsel von einem Typ zu None
(current_section_type_meta is not None and item_section_type is None) or
# Wechsel zwischen verschiedenen Typen
(current_section_type_meta is not None and item_section_type is not None
and current_section_type_meta != item_section_type)
)
)
if is_section_type_change_step2:
# WP-26 v1.1: Section-Type-Wechsel erzwingt Split
_emit(current_chunk_text, current_meta["title"], current_meta["path"],
current_meta["section_type"], current_meta["block_id"])
current_chunk_text = ""
# Reset Meta für nächsten Chunk
current_meta["title"] = item["meta"].section_title
current_meta["path"] = item["meta"].section_path
current_meta["section_type"] = item_section_type
current_meta["block_id"] = item.get("block_id")
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.
WP-26 v1.1: Section-Type-Wechsel erzwingt IMMER einen neuen Chunk.
"""
target = config.get("target", 400); max_tokens = config.get("max", 600)
chunks: List[Chunk] = []; buf: List[RawBlock] = []
current_section_type = None # WP-26 v1.1: Tracking des aktuellen section_type
def _flush_buffer():
"""Hilfsfunktion zum Flushen des Buffers."""
nonlocal buf, current_section_type
if not buf:
return
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=sum(estimate_tokens(x.text) for x in buf),
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 = []
current_section_type = effective_section_type
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
block_section_type = b.section_type
# WP-26 v1.1: Prüfe auf Section-Type-Wechsel
is_section_type_change = (
buf and # Es gibt bereits Blöcke im Buffer
block_section_type is not None and # Neuer Block hat expliziten section_type
current_section_type != block_section_type # Typ hat sich geändert
)
# Flush wenn: Token-Limit überschritten ODER Section-Type-Wechsel
if (curr_tokens + b_tokens > max_tokens and buf) or is_section_type_change:
_flush_buffer()
buf.append(b)
# Update section_type wenn Block einen hat und wir noch keinen haben
if block_section_type and not current_section_type:
current_section_type = block_section_type
# Letzten Buffer flushen
_flush_buffer()
return chunks