Aktualisierung des Chunking-Parsers zur Einführung der Funktion parse_edges_robust zur Extraktion von Kanten-Kandidaten aus Wikilinks und Callouts. Verbesserung der Satzverarbeitung durch die Implementierung der Funktion split_sentences. Aktualisierung der Sektions-Chunking-Strategie auf Version 3.9.6 mit optimierter Handhabung von leeren Überschriften und Carry-Over Logik zur besseren Chunk-Erstellung.
This commit is contained in:
parent
3c5c567077
commit
06fc42ed37
|
|
@ -1,13 +1,23 @@
|
||||||
"""
|
"""
|
||||||
FILE: app/core/chunking/chunking_parser.py
|
FILE: app/core/chunking/chunking_parser.py
|
||||||
DESCRIPTION: Zerlegt Markdown in atomare Blöcke. Hält H1-Überschriften im Stream
|
DESCRIPTION: Zerlegt Markdown in atomare Blöcke. Hält H1-Überschriften im Stream
|
||||||
und gewährleistet die strukturelle Integrität von Callouts.
|
und extrahiert Kanten-Kandidaten (parse_edges_robust).
|
||||||
"""
|
"""
|
||||||
import re
|
import re
|
||||||
from typing import List, Tuple, Set
|
from typing import List, Tuple, Set
|
||||||
from .chunking_models import RawBlock
|
from .chunking_models import RawBlock
|
||||||
from .chunking_utils import extract_frontmatter_from_text
|
from .chunking_utils import extract_frontmatter_from_text
|
||||||
|
|
||||||
|
_WS = re.compile(r'\s+')
|
||||||
|
_SENT_SPLIT = re.compile(r'(?<=[.!?])\s+(?=[A-ZÄÖÜ0-9„(])')
|
||||||
|
|
||||||
|
def split_sentences(text: str) -> list[str]:
|
||||||
|
"""Teilt Text in Sätze auf unter Berücksichtigung deutscher Interpunktion."""
|
||||||
|
text = _WS.sub(' ', text.strip())
|
||||||
|
if not text: return []
|
||||||
|
# Splittet bei Punkt, Ausrufezeichen oder Fragezeichen, gefolgt von Leerzeichen und Großbuchstabe
|
||||||
|
return [p.strip() for p in _SENT_SPLIT.split(text) if p.strip()]
|
||||||
|
|
||||||
def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
||||||
"""Zerlegt Text in logische Einheiten (RawBlocks), inklusive H1-H6."""
|
"""Zerlegt Text in logische Einheiten (RawBlocks), inklusive H1-H6."""
|
||||||
blocks = []
|
blocks = []
|
||||||
|
|
@ -16,17 +26,18 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
||||||
|
|
||||||
# H1 für Metadaten extrahieren
|
# H1 für Metadaten extrahieren
|
||||||
h1_match = re.search(r'^#\s+(.*)', text_without_fm, re.MULTILINE)
|
h1_match = re.search(r'^#\s+(.*)', text_without_fm, re.MULTILINE)
|
||||||
if h1_match: h1_title = h1_match.group(1).strip()
|
if h1_match:
|
||||||
|
h1_title = h1_match.group(1).strip()
|
||||||
|
|
||||||
lines = text_without_fm.split('\n')
|
lines = text_without_fm.split('\n')
|
||||||
buffer = []
|
buffer = []
|
||||||
|
|
||||||
for line in lines:
|
for line in lines:
|
||||||
stripped = line.strip()
|
stripped = line.strip()
|
||||||
heading_match = re.match(r'^(#{1,6})\s+(.*)', stripped)
|
|
||||||
|
|
||||||
|
# Heading-Erkennung (H1 bis H6)
|
||||||
|
heading_match = re.match(r'^(#{1,6})\s+(.*)', stripped)
|
||||||
if heading_match:
|
if heading_match:
|
||||||
# Vorherigen Text-Block abschließen
|
|
||||||
if buffer:
|
if buffer:
|
||||||
content = "\n".join(buffer).strip()
|
content = "\n".join(buffer).strip()
|
||||||
if content:
|
if content:
|
||||||
|
|
@ -59,13 +70,37 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
||||||
|
|
||||||
if buffer:
|
if buffer:
|
||||||
content = "\n".join(buffer).strip()
|
content = "\n".join(buffer).strip()
|
||||||
if content: blocks.append(RawBlock("paragraph", content, None, section_path, current_section_title))
|
if content:
|
||||||
|
blocks.append(RawBlock("paragraph", content, None, section_path, current_section_title))
|
||||||
|
|
||||||
return blocks, h1_title
|
return blocks, h1_title
|
||||||
|
|
||||||
def split_sentences(text: str) -> list[str]:
|
def parse_edges_robust(text: str) -> Set[str]:
|
||||||
"""Teilt Text in Sätze auf unter Berücksichtigung deutscher Interpunktion."""
|
"""Extrahiert Kanten-Kandidaten aus Wikilinks und Callouts."""
|
||||||
text = re.sub(r'\s+', ' ', text.strip())
|
found_edges = set()
|
||||||
if not text: return []
|
# 1. Wikilinks [[rel:kind|target]]
|
||||||
# Splittet bei Satzzeichen, gefolgt von Leerzeichen und Großbuchstaben
|
inlines = re.findall(r'\[\[rel:([^\|\]]+)\|?([^\]]*)\]\]', text)
|
||||||
return [s.strip() for s in re.split(r'(?<=[.!?])\s+(?=[A-ZÄÖÜ0-9„(])', text) if s.strip()]
|
for kind, target in inlines:
|
||||||
|
k = kind.strip().lower()
|
||||||
|
t = target.strip()
|
||||||
|
if k and t: found_edges.add(f"{k}:{t}")
|
||||||
|
|
||||||
|
# 2. Callout Edges > [!edge] kind
|
||||||
|
lines = text.split('\n')
|
||||||
|
current_edge_type = None
|
||||||
|
for line in lines:
|
||||||
|
stripped = line.strip()
|
||||||
|
callout_match = re.match(r'>\s*\[!edge\]\s*([^:\s]+)', stripped)
|
||||||
|
if callout_match:
|
||||||
|
current_edge_type = callout_match.group(1).strip().lower()
|
||||||
|
links = re.findall(r'\[\[([^\]]+)\]\]', stripped)
|
||||||
|
for l in links:
|
||||||
|
if "rel:" not in l: found_edges.add(f"{current_edge_type}:{l}")
|
||||||
|
continue
|
||||||
|
if current_edge_type and stripped.startswith('>'):
|
||||||
|
links = re.findall(r'\[\[([^\]]+)\]\]', stripped)
|
||||||
|
for l in links:
|
||||||
|
if "rel:" not in l: found_edges.add(f"{current_edge_type}:{l}")
|
||||||
|
elif not stripped.startswith('>'):
|
||||||
|
current_edge_type = None
|
||||||
|
return found_edges
|
||||||
|
|
@ -1,7 +1,11 @@
|
||||||
"""
|
"""
|
||||||
FILE: app/core/chunking/chunking_strategies.py
|
FILE: app/core/chunking/chunking_strategies.py
|
||||||
DESCRIPTION: Strategie für atomares Sektions-Chunking v3.9.5.
|
DESCRIPTION: Strategien für atomares Sektions-Chunking v3.9.6.
|
||||||
Implementiert das 'Pack-and-Carry-Over' Verfahren nach Nutzerwunsch.
|
Implementiert das 'Pack-and-Carry-Over' Verfahren:
|
||||||
|
1. Packt ganze Abschnitte basierend auf Schätzung.
|
||||||
|
2. Kein physischer Overflow-Check während des Packens.
|
||||||
|
3. Smart-Zerlegung von Übergrößen mit Carry-Over in die Queue.
|
||||||
|
- Hard-Split-Logik für strict_heading_split integriert.
|
||||||
"""
|
"""
|
||||||
from typing import List, Dict, Any, Optional
|
from typing import List, Dict, Any, Optional
|
||||||
from .chunking_models import RawBlock, Chunk
|
from .chunking_models import RawBlock, Chunk
|
||||||
|
|
@ -16,10 +20,7 @@ def _create_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
|
||||||
|
|
||||||
def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, doc_title: str = "") -> List[Chunk]:
|
def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, doc_title: str = "") -> List[Chunk]:
|
||||||
"""
|
"""
|
||||||
Universelle Sektions-Strategie:
|
Universelle Heading-Strategie mit Carry-Over Logik.
|
||||||
- Smart-Edge=True: Packt Sektionen basierend auf Schätzung (Regel 1-3).
|
|
||||||
- Smart-Edge=False: Hard Split an Überschriften (außer leere Header).
|
|
||||||
- Strict=True erzwingt Hard Split Verhalten innerhalb der Smart-Logik.
|
|
||||||
"""
|
"""
|
||||||
smart_edge = config.get("enable_smart_edge_allocation", True)
|
smart_edge = config.get("enable_smart_edge_allocation", True)
|
||||||
strict = config.get("strict_heading_split", False)
|
strict = config.get("strict_heading_split", False)
|
||||||
|
|
@ -46,66 +47,73 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
|
||||||
for b in blocks:
|
for b in blocks:
|
||||||
if b.kind == "heading" and b.level <= split_level:
|
if b.kind == "heading" and b.level <= split_level:
|
||||||
if curr_blocks:
|
if curr_blocks:
|
||||||
sections.append({"text": "\n\n".join([x.text for x in curr_blocks]),
|
sections.append({
|
||||||
"meta": curr_blocks[0], "is_empty": len(curr_blocks) == 1})
|
"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"
|
||||||
|
})
|
||||||
curr_blocks = [b]
|
curr_blocks = [b]
|
||||||
else: curr_blocks.append(b)
|
else:
|
||||||
|
curr_sec_has_content = True
|
||||||
|
curr_blocks.append(b)
|
||||||
if curr_blocks:
|
if curr_blocks:
|
||||||
sections.append({"text": "\n\n".join([x.text for x in curr_blocks]),
|
sections.append({
|
||||||
"meta": curr_blocks[0], "is_empty": len(curr_blocks) == 1})
|
"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"
|
||||||
|
})
|
||||||
|
|
||||||
# --- SCHRITT 2: Verarbeitung der Queue ---
|
# --- SCHRITT 2: Verarbeitung der Queue ---
|
||||||
queue = list(sections)
|
queue = list(sections)
|
||||||
current_chunk_text = ""
|
current_chunk_text = ""
|
||||||
current_meta = {"title": None, "path": "/"}
|
current_meta = {"title": None, "path": "/"}
|
||||||
|
|
||||||
# Hard-Split-Bedingung: Entweder Smart-Edge aus ODER Profil ist Strict
|
# Bestimmung des Modus: Hard-Split wenn smart_edge=False ODER strict=True
|
||||||
is_hard_split_mode = (not smart_edge) or (strict)
|
is_hard_split_mode = (not smart_edge) or (strict)
|
||||||
|
|
||||||
while queue:
|
while queue:
|
||||||
item = queue.pop(0)
|
item = queue.pop(0)
|
||||||
item_text = item["text"]
|
item_text = item["text"]
|
||||||
|
|
||||||
# Initialisierung für neuen Chunk
|
|
||||||
if not current_chunk_text:
|
if not current_chunk_text:
|
||||||
current_meta["title"] = item["meta"].section_title
|
current_meta["title"] = item["meta"].section_title
|
||||||
current_meta["path"] = item["meta"].section_path
|
current_meta["path"] = item["meta"].section_path
|
||||||
|
|
||||||
# FALL A: Hard Split Modus (Regel: Trenne bei jeder Sektion <= Level)
|
# FALL A: HARD SPLIT MODUS
|
||||||
if is_hard_split_mode:
|
if is_hard_split_mode:
|
||||||
# Regel: Leere Überschriften verbleiben am nächsten Chunk
|
# Leere Überschriften (H1 vor H2) werden mit dem nächsten Item verschmolzen
|
||||||
if item.get("is_empty", False) and queue:
|
if item.get("is_empty", False) and queue:
|
||||||
current_chunk_text = (current_chunk_text + "\n\n" + item_text).strip()
|
current_chunk_text = (current_chunk_text + "\n\n" + item_text).strip()
|
||||||
continue # Nimm das nächste Item dazu
|
continue
|
||||||
|
|
||||||
combined = (current_chunk_text + "\n\n" + item_text).strip()
|
combined = (current_chunk_text + "\n\n" + item_text).strip()
|
||||||
|
# Wenn durch das Verschmelzen das Limit gesprengt würde, flashen wir vorher
|
||||||
if estimate_tokens(combined) > max_tokens and current_chunk_text:
|
if estimate_tokens(combined) > max_tokens and current_chunk_text:
|
||||||
# Falls es trotz Hard-Split zu groß wird, flashen wir erst den alten Teil
|
|
||||||
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
||||||
current_chunk_text = item_text
|
current_chunk_text = item_text
|
||||||
else:
|
else:
|
||||||
current_chunk_text = combined
|
current_chunk_text = combined
|
||||||
|
|
||||||
# Im Hard Split flashen wir nach jeder Sektion, die nicht leer ist
|
# Im Hard-Split wird nach jeder nicht-leeren Sektion geflasht
|
||||||
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
||||||
current_chunk_text = ""
|
current_chunk_text = ""
|
||||||
continue
|
continue
|
||||||
|
|
||||||
# FALL B: Smart Mode (Regel 1-3)
|
# 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_text = (current_chunk_text + "\n\n" + item_text).strip() if current_chunk_text else item_text
|
||||||
combined_est = estimate_tokens(combined_text)
|
combined_est = estimate_tokens(combined_text)
|
||||||
|
|
||||||
if combined_est <= max_tokens:
|
if combined_est <= max_tokens:
|
||||||
# Regel 1 & 2: Passt nach Schätzung -> Aufnehmen
|
# Regel 1 & 2: Passt rein -> Aufnehmen
|
||||||
current_chunk_text = combined_text
|
current_chunk_text = combined_text
|
||||||
else:
|
else:
|
||||||
# Regel 3: Passt nicht -> Entweder Puffer flashen oder Item zerlegen
|
|
||||||
if current_chunk_text:
|
if current_chunk_text:
|
||||||
|
# Regel 2: Flashen an Sektionsgrenze, Item zurücklegen
|
||||||
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
||||||
current_chunk_text = ""
|
current_chunk_text = ""
|
||||||
queue.insert(0, item) # Item für neuen Chunk zurücklegen
|
queue.insert(0, item)
|
||||||
else:
|
else:
|
||||||
# Einzelne Sektion zu groß -> Smart Zerlegung
|
# Regel 3: Einzelne Sektion zu groß -> Smart Zerlegung
|
||||||
sents = split_sentences(item_text)
|
sents = split_sentences(item_text)
|
||||||
header_prefix = item["meta"].text if item["meta"].kind == "heading" else ""
|
header_prefix = item["meta"].text if item["meta"].kind == "heading" else ""
|
||||||
|
|
||||||
|
|
@ -119,11 +127,11 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
|
||||||
|
|
||||||
_emit(" ".join(take_sents), current_meta["title"], current_meta["path"])
|
_emit(" ".join(take_sents), current_meta["title"], current_meta["path"])
|
||||||
|
|
||||||
# Carry-Over: Rest wird vorne in die Queue geschoben
|
|
||||||
if sents:
|
if sents:
|
||||||
remainder = " ".join(sents)
|
remainder = " ".join(sents)
|
||||||
if header_prefix and not remainder.startswith(header_prefix):
|
if header_prefix and not remainder.startswith(header_prefix):
|
||||||
remainder = header_prefix + "\n\n" + remainder
|
remainder = header_prefix + "\n\n" + remainder
|
||||||
|
# Carry-Over: Rest wird vorne in die Queue geschoben
|
||||||
queue.insert(0, {"text": remainder, "meta": item["meta"], "is_split": True})
|
queue.insert(0, {"text": remainder, "meta": item["meta"], "is_split": True})
|
||||||
|
|
||||||
if current_chunk_text:
|
if current_chunk_text:
|
||||||
|
|
@ -132,7 +140,7 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
|
||||||
return chunks
|
return chunks
|
||||||
|
|
||||||
def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, doc_title: str = "") -> List[Chunk]:
|
def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, doc_title: str = "") -> List[Chunk]:
|
||||||
"""Standard Sliding Window Strategie."""
|
"""Basis-Sliding-Window für flache Texte."""
|
||||||
target = config.get("target", 400); max_tokens = config.get("max", 600)
|
target = config.get("target", 400); max_tokens = config.get("max", 600)
|
||||||
chunks: List[Chunk] = []; buf: List[RawBlock] = []
|
chunks: List[Chunk] = []; buf: List[RawBlock] = []
|
||||||
for b in blocks:
|
for b in blocks:
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue
Block a user