Optimierung des Chunking-Parsers zur Unterstützung atomarer Blöcke und Verbesserung der Satzverarbeitung. Aktualisierung der Sektions-Chunking-Strategie auf Version 3.9.0 mit regelkonformer Implementierung und Anpassungen an der Warteschlangen-Verarbeitung für Carry-Over. Verbesserte Handhabung von Überschriften und Metadaten zur Gewährleistung der strukturellen Integrität.
This commit is contained in:
parent
6b83879741
commit
8f65e550c8
|
|
@ -1,6 +1,6 @@
|
|||
"""
|
||||
FILE: app/core/chunking/chunking_parser.py
|
||||
DESCRIPTION: Zerlegt Markdown in logische Blöcke. Hält H1-Überschriften im Stream
|
||||
DESCRIPTION: Zerlegt Markdown in atomare Blöcke. Hält H1-Überschriften im Stream
|
||||
und gewährleistet die strukturelle Integrität von Callouts.
|
||||
"""
|
||||
import re
|
||||
|
|
@ -8,15 +8,12 @@ from typing import List, Tuple, Set
|
|||
from .chunking_models import RawBlock
|
||||
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())
|
||||
text = re.sub(r'\s+', ' ', text.strip())
|
||||
if not text: return []
|
||||
# Splittet bei Punkt, Ausrufezeichen oder Fragezeichen, gefolgt von Großbuchstabe
|
||||
return [p.strip() for p in _SENT_SPLIT.split(text) if p.strip()]
|
||||
# Splittet bei Satzzeichen, gefolgt von Leerzeichen und Großbuchstaben
|
||||
return [s.strip() for s in re.split(r'(?<=[.!?])\s+(?=[A-ZÄÖÜ0-9„(])', text) if s.strip()]
|
||||
|
||||
def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
||||
"""Zerlegt Text in logische Einheiten (RawBlocks), inklusive H1-H6."""
|
||||
|
|
@ -24,7 +21,6 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
|||
h1_title = "Dokument"; section_path = "/"; current_section_title = None
|
||||
fm, text_without_fm = extract_frontmatter_from_text(md_text)
|
||||
|
||||
# H1 für Note-Titel extrahieren
|
||||
h1_match = re.search(r'^#\s+(.*)', text_without_fm, re.MULTILINE)
|
||||
if h1_match: h1_title = h1_match.group(1).strip()
|
||||
|
||||
|
|
@ -33,28 +29,23 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
|||
|
||||
for line in lines:
|
||||
stripped = line.strip()
|
||||
|
||||
# Heading-Erkennung (H1 bis H6)
|
||||
heading_match = re.match(r'^(#{1,6})\s+(.*)', stripped)
|
||||
|
||||
if heading_match:
|
||||
if buffer:
|
||||
content = "\n".join(buffer).strip()
|
||||
if content:
|
||||
blocks.append(RawBlock("paragraph", content, None, section_path, current_section_title))
|
||||
buffer = []
|
||||
|
||||
level = len(heading_match.group(1))
|
||||
title = heading_match.group(2).strip()
|
||||
|
||||
if level == 1:
|
||||
current_section_title = title; section_path = "/"
|
||||
elif level == 2:
|
||||
current_section_title = title; section_path = f"/{current_section_title}"
|
||||
|
||||
blocks.append(RawBlock("heading", stripped, level, section_path, current_section_title))
|
||||
continue
|
||||
|
||||
# Trenner oder Leerzeilen beenden Blöcke, außer innerhalb von Callouts
|
||||
if (not stripped or stripped == "---") and not line.startswith('>'):
|
||||
if buffer:
|
||||
content = "\n".join(buffer).strip()
|
||||
|
|
@ -69,7 +60,6 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
|||
if buffer:
|
||||
content = "\n".join(buffer).strip()
|
||||
if content: blocks.append(RawBlock("paragraph", content, None, section_path, current_section_title))
|
||||
|
||||
return blocks, h1_title
|
||||
|
||||
def parse_edges_robust(text: str) -> Set[str]:
|
||||
|
|
@ -79,7 +69,6 @@ def parse_edges_robust(text: str) -> Set[str]:
|
|||
for kind, target in inlines:
|
||||
k = kind.strip().lower(); t = target.strip()
|
||||
if k and t: found_edges.add(f"{k}:{t}")
|
||||
|
||||
lines = text.split('\n'); current_edge_type = None
|
||||
for line in lines:
|
||||
stripped = line.strip()
|
||||
|
|
|
|||
|
|
@ -1,10 +1,7 @@
|
|||
"""
|
||||
FILE: app/core/chunking/chunking_strategies.py
|
||||
DESCRIPTION: Strategien für atomares Sektions-Chunking v3.8.5.
|
||||
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.
|
||||
DESCRIPTION: Universelle Strategie für atomares Sektions-Chunking v3.9.0.
|
||||
Regelkonforme Implementierung: Pack-Sections, Trust Estimation, Carry-Over.
|
||||
"""
|
||||
from typing import List, Dict, Any, Optional
|
||||
from .chunking_models import RawBlock, Chunk
|
||||
|
|
@ -18,9 +15,6 @@ def _create_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
|
|||
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]:
|
||||
"""
|
||||
Universelle Heading-Strategie mit Fallunterscheidung für Smart-Edge-Allocation.
|
||||
"""
|
||||
smart_edge = config.get("enable_smart_edge_allocation", True)
|
||||
target = config.get("target", 400)
|
||||
max_tokens = config.get("max", 600)
|
||||
|
|
@ -43,10 +37,11 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
|
|||
if not smart_edge:
|
||||
buf = []
|
||||
for b in blocks:
|
||||
# Trenne bei jeder Überschrift <= split_level
|
||||
# Trenne hart bei Überschrift <= split_level
|
||||
if b.kind == "heading" and b.level <= split_level:
|
||||
# Nur flashen, wenn der Puffer nicht nur aus der aktuellen Überschrift besteht
|
||||
if buf and not (len(buf) == 1 and buf[0].kind == "heading"):
|
||||
# Prüfe, ob Puffer mehr als nur Überschriften enthält (keine leeren Chunks)
|
||||
has_content = any(x.kind != "heading" for x in buf)
|
||||
if buf and has_content:
|
||||
_emit("\n\n".join([x.text for x in buf]), buf[0].section_title, buf[0].section_path)
|
||||
buf = []
|
||||
buf.append(b)
|
||||
|
|
@ -54,49 +49,52 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
|
|||
return chunks
|
||||
|
||||
# --- FALL B: SMART EDGE ALLOCATION (Pack-and-Carry-Over) ---
|
||||
# 1. Gruppierung in atomare Sektions-Einheiten (Sektions-Isolation)
|
||||
sections: List[List[RawBlock]] = []
|
||||
curr = []
|
||||
# 1. Gruppierung in atomare Einheiten (Sektions-Isolation)
|
||||
sections: List[Dict[str, Any]] = []
|
||||
curr_blocks = []
|
||||
for b in blocks:
|
||||
if b.kind == "heading" and b.level <= split_level:
|
||||
if curr: sections.append(curr)
|
||||
curr = [b]
|
||||
else: curr.append(b)
|
||||
if curr: sections.append(curr)
|
||||
if curr_blocks:
|
||||
sections.append({"text": "\n\n".join([x.text for x in curr_blocks]), "meta": curr_blocks[0]})
|
||||
curr_blocks = [b]
|
||||
else: curr_blocks.append(b)
|
||||
if curr_blocks:
|
||||
sections.append({"text": "\n\n".join([x.text for x in curr_blocks]), "meta": curr_blocks[0]})
|
||||
|
||||
# 2. Queue-Management für Carry-Over
|
||||
processing_queue = [{"blocks": s, "text": "\n\n".join([b.text for b in s])} for s in sections]
|
||||
# 2. Warteschlangen-Verarbeitung (Regel 1-3)
|
||||
# Wir nutzen eine Liste als Queue für Carry-Over-Reste
|
||||
queue = list(sections)
|
||||
current_chunk_text = ""
|
||||
current_meta = {"title": None, "path": "/"}
|
||||
|
||||
while processing_queue:
|
||||
item = processing_queue.pop(0)
|
||||
while queue:
|
||||
item = queue.pop(0)
|
||||
item_text = item["text"]
|
||||
item_tokens = estimate_tokens(item_text)
|
||||
|
||||
# Metadaten-Initialisierung falls Chunk leer
|
||||
if not current_chunk_text and "blocks" in item:
|
||||
current_meta["title"] = item["blocks"][0].section_title
|
||||
current_meta["path"] = item["blocks"][0].section_path
|
||||
# Initialisiere Metadaten für einen neuen Chunk
|
||||
if not current_chunk_text:
|
||||
current_meta["title"] = item["meta"].section_title
|
||||
current_meta["path"] = item["meta"].section_path
|
||||
|
||||
combined_est = estimate_tokens(current_chunk_text + "\n\n" + item_text) if current_chunk_text else item_tokens
|
||||
# Schätzung (Regel 2: Wir verlassen uns darauf)
|
||||
combined_text = (current_chunk_text + "\n\n" + item_text).strip() if current_chunk_text else item_text
|
||||
combined_est = estimate_tokens(combined_text)
|
||||
|
||||
# Regel 1: Passt die vollständige Sektion nach Schätzung rein? (Kein harter Overflow-Check)
|
||||
if combined_est <= max_tokens:
|
||||
current_chunk_text = (current_chunk_text + "\n\n" + item_text).strip()
|
||||
# Regel 1: Vollständiger Abschnitt passt -> Aufnehmen
|
||||
current_chunk_text = combined_text
|
||||
else:
|
||||
# Regel 2: Wenn Puffer voll -> Emittieren und Sektion zurücklegen
|
||||
# Er passt nicht ganz rein.
|
||||
if current_chunk_text:
|
||||
# Puffer ist bereits gefüllt -> Wegschreiben, Item zurück in die Queue
|
||||
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
||||
current_chunk_text = ""
|
||||
processing_queue.insert(0, item)
|
||||
queue.insert(0, item)
|
||||
else:
|
||||
# Regel 3: Einzelne Sektion zu groß -> Smart Zerlegung
|
||||
# Regel 3: Einzelner Abschnitt allein ist > max -> Smart Zerlegung
|
||||
sents = split_sentences(item_text)
|
||||
header_text = ""
|
||||
if "blocks" in item and item["blocks"][0].kind == "heading":
|
||||
header_text = item["blocks"][0].text
|
||||
|
||||
header_prefix = item["meta"].text if item["meta"].kind == "heading" else ""
|
||||
|
||||
take_sents = []; take_len = 0
|
||||
while sents:
|
||||
s = sents.pop(0)
|
||||
|
|
@ -105,39 +103,38 @@ def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id:
|
|||
sents.insert(0, s); break
|
||||
take_sents.append(s); take_len += slen
|
||||
|
||||
# Ersten Teil emittieren
|
||||
_emit(" ".join(take_sents), current_meta["title"], current_meta["path"])
|
||||
|
||||
# Carry-Over: Der Rest wird an den Anfang der Queue geschoben
|
||||
# Rest als Carry-Over zurück in die Queue (Regel 3)
|
||||
if sents:
|
||||
remainder_text = " ".join(sents)
|
||||
# Falls wir einen Header haben, fügen wir ihn dem Rest für den Kontext hinzu
|
||||
if header_text and not remainder_text.startswith(header_text):
|
||||
remainder_text = header_text + "\n\n" + remainder_text
|
||||
processing_queue.insert(0, {"text": remainder_text, "is_split": True})
|
||||
# Kontext-Erhalt: Überschrift für den Rest wiederholen
|
||||
if header_prefix and not remainder_text.startswith(header_prefix):
|
||||
remainder_text = header_prefix + "\n\n" + remainder_text
|
||||
queue.insert(0, {"text": remainder_text, "meta": item["meta"]})
|
||||
|
||||
# Letzten Rest wegschreiben
|
||||
if current_chunk_text:
|
||||
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
||||
|
||||
return chunks
|
||||
|
||||
def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, context_prefix: str = "") -> List[Chunk]:
|
||||
"""Sliding Window Strategie: Erhalten für alternative Anwendungsfälle."""
|
||||
"""Sliding Window: Unverändert erhalten für Standard-Typen."""
|
||||
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)
|
||||
current_tokens = sum(estimate_tokens(x.text) for x in buf) if buf else 0
|
||||
if current_tokens + b_tokens > max_tokens and buf:
|
||||
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 = 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))
|
||||
win = _create_win(doc_title=context_prefix, sec_title=buf[0].section_title, text=txt)
|
||||
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))
|
||||
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
|
||||
win = _create_win(doc_title=context_prefix, sec_title=buf[0].section_title, text=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=buf[0].section_title, section_path=buf[0].section_path, neighbors_prev=None, neighbors_next=None))
|
||||
|
||||
return chunks
|
||||
Loading…
Reference in New Issue
Block a user