Lars/mindnet
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 21 Wiki Activity

Aktualisierung des Chunking-Parsers zur Verbesserung der Blockverarbeitung und Beschreibung. Anpassungen an der atomaren Sektions-Chunking-Strategie mit optimierter Token-Schätzung und neuen Hilfsfunktionen zur besseren Handhabung von großen Sektionen.

Browse Source
This commit is contained in:
Lars 2025-12-29 21:37:11 +01:00
parent b1a897e51c
commit 96b4f65cd1
2 changed files with 55 additions and 68 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
app/core/chunking/chunking_parser.py
View File

@ -1,7 +1,7 @@
""" """
FILE: app/core/chunking/chunking_parser.py FILE: app/core/chunking/chunking_parser.py
DESCRIPTION: Zerlegt Markdown in logische Blöcke. Hält H1-Überschriften im Stream DESCRIPTION: Zerlegt Markdown in logische Einheiten (RawBlocks).
und optimiert die Block-Trennung für atomares Sektions-Chunking. Hält H1-Überschriften im Inhalts-Stream.
""" """
import re import re
from typing import List, Tuple, Set from typing import List, Tuple, Set
@ -15,7 +15,6 @@ def split_sentences(text: str) -> list[str]:
"""Teilt Text in Sätze auf unter Berücksichtigung deutscher Interpunktion.""" """Teilt Text in Sätze auf unter Berücksichtigung deutscher Interpunktion."""
text = _WS.sub(' ', text.strip()) text = _WS.sub(' ', text.strip())
if not text: return [] if not text: return []
# Splittet bei Satzzeichen, gefolgt von Leerzeichen und Großbuchstaben
return [p.strip() for p in _SENT_SPLIT.split(text) if p.strip()] 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]:
@ -28,7 +27,7 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
# Frontmatter entfernen # Frontmatter entfernen
fm, text_without_fm = extract_frontmatter_from_text(md_text) fm, text_without_fm = extract_frontmatter_from_text(md_text)
# H1 für Note-Metadaten extrahieren # H1 für Note-Titel 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: if h1_match:
h1_title = h1_match.group(1).strip() h1_title = h1_match.group(1).strip()
@ -42,7 +41,7 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
# Heading-Erkennung (H1 bis H6) # Heading-Erkennung (H1 bis H6)
heading_match = re.match(r'^(#{1,6})\s+(.*)', stripped) heading_match = re.match(r'^(#{1,6})\s+(.*)', stripped)
if heading_match: if heading_match:
# Vorherigen Block abschließen # Vorherigen Text-Block abschließen
if buffer: if buffer:
content = "\n".join(buffer).strip() content = "\n".join(buffer).strip()
if content: if content:
@ -52,7 +51,7 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
level = len(heading_match.group(1)) level = len(heading_match.group(1))
title = heading_match.group(2).strip() title = heading_match.group(2).strip()
# Update der Pfad-Metadaten für die folgenden Blöcke # Pfad- und Titel-Update für die Metadaten der folgenden Blöcke
if level == 1: if level == 1:
current_section_title = title current_section_title = title
section_path = "/" section_path = "/"
@ -60,11 +59,11 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
current_section_title = title current_section_title = title
section_path = f"/{current_section_title}" section_path = f"/{current_section_title}"
# Die Überschrift als regulären Block hinzufügen (Fix: H1 bleibt im Text) # Die Überschrift selbst als Block hinzufügen
blocks.append(RawBlock("heading", stripped, level, section_path, current_section_title)) blocks.append(RawBlock("heading", stripped, level, section_path, current_section_title))
continue continue
# Trenner (---) oder Leerzeilen beenden Blöcke, außer in Callouts # Trenner oder Leerzeilen beenden Blöcke, außer innerhalb von Callouts
if (not stripped or stripped == "---") and not line.startswith('>'): if (not stripped or stripped == "---") and not line.startswith('>'):
if buffer: if buffer:
content = "\n".join(buffer).strip() content = "\n".join(buffer).strip()
@ -76,7 +75,6 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
else: else:
buffer.append(line) buffer.append(line)
# Letzten Puffer leeren
if buffer: if buffer:
content = "\n".join(buffer).strip() content = "\n".join(buffer).strip()
if content: if content:
@ -87,14 +85,12 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
def parse_edges_robust(text: str) -> Set[str]: def parse_edges_robust(text: str) -> Set[str]:
"""Extrahiert Kanten-Kandidaten aus Wikilinks und Callouts.""" """Extrahiert Kanten-Kandidaten aus Wikilinks und Callouts."""
found_edges = set() found_edges = set()
# 1. Wikilinks [[rel:kind|target]]
inlines = re.findall(r'\[\[rel:([^\|\]]+)\|?([^\]]*)\]\]', text) inlines = re.findall(r'\[\[rel:([^\|\]]+)\|?([^\]]*)\]\]', text)
for kind, target in inlines: for kind, target in inlines:
k = kind.strip().lower() k = kind.strip().lower()
t = target.strip() t = target.strip()
if k and t: found_edges.add(f"{k}:{t}") if k and t: found_edges.add(f"{k}:{t}")
# 2. Callout Edges > [!edge] kind
lines = text.split('\n') lines = text.split('\n')
current_edge_type = None current_edge_type = None
for line in lines: for line in lines:

107
app/core/chunking/chunking_strategies.py
View File

@ -1,20 +1,15 @@
""" """
FILE: app/core/chunking/chunking_strategies.py FILE: app/core/chunking/chunking_strategies.py
DESCRIPTION: Strategien für atomares Sektions-Chunking v3.4.1. DESCRIPTION: Universelle Strategie für atomares Sektions-Chunking v3.5.0.
Garantiert Sektions-Integrität (Atomic Units) durch Look-Ahead. Garantiert Sektions-Integrität durch präventives Chunk-Management.
""" """
import math
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
from .chunking_utils import estimate_tokens from .chunking_utils import estimate_tokens
from .chunking_parser import split_sentences 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: def _create_context_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
"""Baut den Breadcrumb-Kontext für das Embedding-Fenster (H1 > H2).""" """Baut den Breadcrumb-Kontext für das Embedding-Fenster."""
parts = [] parts = []
if doc_title: parts.append(doc_title) if doc_title: parts.append(doc_title)
if sec_title and sec_title != doc_title: parts.append(sec_title) if sec_title and sec_title != doc_title: parts.append(sec_title)
@ -23,10 +18,9 @@ def _create_context_win(doc_title: str, sec_title: Optional[str], text: 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]:
""" """
Sektions-Chunking: Behandelt Abschnitte als unteilbare Einheiten. Universelles Sektions-Chunking: Packt Sektionen in Chunks.
Schiebt ganze Abschnitte in den nächsten Chunk, falls das Limit erreicht ist. Bei Überlauf wird die komplette Sektion in den nächsten Chunk geschoben.
""" """
strict = config.get("strict_heading_split", False)
target = config.get("target", 400) target = config.get("target", 400)
max_tokens = config.get("max", 600) max_tokens = config.get("max", 600)
split_level = config.get("split_level", 2) split_level = config.get("split_level", 2)
@ -35,24 +29,23 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
chunks: List[Chunk] = [] chunks: List[Chunk] = []
def _emit_chunk(block_list: List[RawBlock]): # --- HILFSFUNKTION: Erzeugt einen Chunk aus einer Blockliste ---
"""Erzeugt ein finales Chunk-Objekt aus einer Liste von Blöcken.""" def _create_chunk_from_blocks(block_list: List[RawBlock]):
if not block_list: return if not block_list: return
txt = "\n\n".join([b.text for b in block_list]) txt = "\n\n".join([b.text for b in block_list])
idx = len(chunks) idx = len(chunks)
# Metadaten vom ersten Block der Gruppe (Header)
title = block_list[0].section_title title = block_list[0].section_title
path = block_list[0].section_path path = block_list[0].section_path
win = _create_context_win(doc_title, title, txt) win = _create_context_win(doc_title, title, txt)
chunks.append(Chunk( chunks.append(Chunk(
id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx, id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx,
text=txt, window=win, token_count=_safe_estimate_tokens(txt), text=txt, window=win, token_count=estimate_tokens(txt),
section_title=title, section_path=path, section_title=title, section_path=path,
neighbors_prev=None, neighbors_next=None neighbors_prev=None, neighbors_next=None
)) ))
def _emit_split_section(sec_blocks: List[RawBlock]): # --- HILFSFUNKTION: Splittet eine einzelne Sektion, die > max ist ---
"""Splittet eine einzelne Sektion, die für sich allein zu groß ist.""" def _split_giant_section(sec_blocks: List[RawBlock]):
full_text = "\n\n".join([b.text for b in sec_blocks]) full_text = "\n\n".join([b.text for b in sec_blocks])
main_title = sec_blocks[0].section_title main_title = sec_blocks[0].section_title
main_path = sec_blocks[0].section_path main_path = sec_blocks[0].section_path
@ -62,15 +55,17 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
cur_sents = []; sub_len = 0 cur_sents = []; sub_len = 0
for s in sents: for s in sents:
slen = _safe_estimate_tokens(s) slen = estimate_tokens(s)
if sub_len + slen > target and cur_sents: if sub_len + slen > target and cur_sents:
_emit_chunk([RawBlock("paragraph", " ".join(cur_sents), None, main_path, main_title)]) combined_text = " ".join(cur_sents)
# Header Injection für den Kontext im nächsten Teil-Chunk _create_chunk_from_blocks([RawBlock("paragraph", combined_text, None, main_path, main_title)])
# Context Injection: Überschrift für den nächsten Teil-Chunk
ov_s = [header_text] if header_text else [] ov_s = [header_text] if header_text else []
ov_l = _safe_estimate_tokens(header_text) if header_text else 0 ov_l = estimate_tokens(header_text) if header_text else 0
for os in reversed(cur_sents): for os in reversed(cur_sents):
if os == header_text: continue if os == header_text: continue
t_len = _safe_estimate_tokens(os) t_len = estimate_tokens(os)
if ov_l + t_len < overlap: if ov_l + t_len < overlap:
ov_s.insert(len(ov_s)-1 if header_text else 0, os) ov_s.insert(len(ov_s)-1 if header_text else 0, os)
ov_l += t_len ov_l += t_len
@ -78,13 +73,15 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
cur_sents = list(ov_s); cur_sents.append(s); sub_len = ov_l + slen cur_sents = list(ov_s); cur_sents.append(s); sub_len = ov_l + slen
else: else:
cur_sents.append(s); sub_len += slen 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) if cur_sents:
_create_chunk_from_blocks([RawBlock("paragraph", " ".join(cur_sents), None, main_path, main_title)])
# 1. SCHRITT: Gruppierung in atomare Sektions-Einheiten
sections: List[List[RawBlock]] = [] sections: List[List[RawBlock]] = []
curr_sec: List[RawBlock] = [] curr_sec: List[RawBlock] = []
for b in blocks: for b in blocks:
# Eine neue Überschrift auf oder unter dem split_level startet eine neue Sektion
if b.kind == "heading" and b.level <= split_level: if b.kind == "heading" and b.level <= split_level:
if curr_sec: sections.append(curr_sec) if curr_sec: sections.append(curr_sec)
curr_sec = [b] curr_sec = [b]
@ -92,54 +89,48 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
curr_sec.append(b) curr_sec.append(b)
if curr_sec: sections.append(curr_sec) if curr_sec: sections.append(curr_sec)
# SCHRITT 2: Verarbeitung der Sektionen mit strengem Look-Ahead # 2. SCHRITT: Sektionen in Chunks packen (Das universelle Pack-Verfahren)
current_chunk_buf = [] candidate_chunk: List[RawBlock] = []
current_tokens = 0 candidate_tokens = 0
for sec in sections: for sec in sections:
sec_text = "\n\n".join([b.text for b in sec]) sec_text = "\n\n".join([b.text for b in sec])
sec_tokens = _safe_estimate_tokens(sec_text) sec_tokens = estimate_tokens(sec_text)
if current_chunk_buf: # Passt diese gesamte Sektion noch in den laufenden Chunk?
# PRÜFUNG: Würde die neue Sektion den aktuellen Chunk sprengen? if candidate_tokens + sec_tokens <= max_tokens:
# ODER: Haben wir das Target bereits erreicht und fangen lieber neu an? candidate_chunk.extend(sec)
if (current_tokens + sec_tokens > max_tokens) or (current_tokens >= target): candidate_tokens = estimate_tokens("\n\n".join([b.text for b in candidate_chunk]))
_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: else:
current_chunk_buf.extend(sec) # Falls der aktuelle Chunk nicht leer ist: Raus damit, bevor die neue Sektion kommt
current_tokens += sec_tokens + 2 # +2 für Newline Join if candidate_chunk:
_create_chunk_from_blocks(candidate_chunk)
candidate_chunk = []
candidate_tokens = 0
# Letzten Puffer schreiben # Die neue Sektion ist nun allein. Ist sie selbst zu groß?
if current_chunk_buf: if sec_tokens > max_tokens:
_emit_chunk(current_chunk_buf) _split_giant_section(sec)
else:
candidate_chunk = list(sec)
candidate_tokens = sec_tokens
# Letzten Rest wegschreiben
if candidate_chunk:
_create_chunk_from_blocks(candidate_chunk)
return chunks return chunks
def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, context_prefix: str = "") -> List[Chunk]: 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.""" """Standard-Sliding-Window für flache Texte ohne Sektionsfokus."""
target = config.get("target", 400) target = config.get("target", 400)
max_tokens = config.get("max", 600) max_tokens = config.get("max", 600)
chunks: List[Chunk] = [] chunks: List[Chunk] = []
buf: List[RawBlock] = [] buf: List[RawBlock] = []
for b in blocks: for b in blocks:
b_tokens = _safe_estimate_tokens(b.text) b_tokens = estimate_tokens(b.text)
current_tokens = sum(_safe_estimate_tokens(x.text) for x in buf) if buf else 0 current_tokens = sum(estimate_tokens(x.text) for x in buf) if buf else 0
if current_tokens + b_tokens > max_tokens and buf: if current_tokens + b_tokens > max_tokens and buf:
txt = "\n\n".join([x.text for x in buf]) txt = "\n\n".join([x.text for x in buf])
idx = len(chunks) idx = len(chunks)
@ -150,5 +141,5 @@ def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note
if buf: if buf:
txt = "\n\n".join([x.text for x in buf]); idx = len(chunks) 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 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)) 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))
return chunks return chunks