WP4d #16
|
|
@ -1,6 +1,6 @@
|
|||
"""
|
||||
FILE: app/core/chunking/chunking_parser.py
|
||||
DESCRIPTION: Zerlegt Markdown in atomare Blöcke. Hält H1-Überschriften im Stream
|
||||
DESCRIPTION: Zerlegt Markdown in logische Blöcke. Hält H1-Überschriften im Stream
|
||||
und gewährleistet die strukturelle Integrität von Callouts.
|
||||
"""
|
||||
import re
|
||||
|
|
@ -8,28 +8,25 @@ 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 = re.sub(r'\s+', ' ', text.strip())
|
||||
text = _WS.sub(' ', text.strip())
|
||||
if not text: return []
|
||||
# Splittet bei Satzzeichen, gefolgt von Leerzeichen und Großbuchstaben
|
||||
sentences = re.split(r'(?<=[.!?])\s+(?=[A-ZÄÖÜ0-9„(])', text)
|
||||
return [s.strip() for s in sentences if s.strip()]
|
||||
# Splittet bei Punkt, Ausrufezeichen oder Fragezeichen, gefolgt von Großbuchstabe
|
||||
return [p.strip() for p in _SENT_SPLIT.split(text) if p.strip()]
|
||||
|
||||
def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
||||
"""Zerlegt Text in logische Einheiten (RawBlocks), inklusive H1-H6."""
|
||||
blocks = []
|
||||
h1_title = "Dokument"
|
||||
section_path = "/"
|
||||
current_section_title = None
|
||||
|
||||
# Frontmatter entfernen
|
||||
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()
|
||||
if h1_match: h1_title = h1_match.group(1).strip()
|
||||
|
||||
lines = text_without_fm.split('\n')
|
||||
buffer = []
|
||||
|
|
@ -57,18 +54,15 @@ def parse_blocks(md_text: str) -> Tuple[List[RawBlock], str]:
|
|||
blocks.append(RawBlock("heading", stripped, level, section_path, current_section_title))
|
||||
continue
|
||||
|
||||
# Trenner (---) beenden Blöcke, Leerzeilen nur wenn nicht in Callout
|
||||
if stripped == "---" and not line.startswith('>'):
|
||||
# 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()
|
||||
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))
|
||||
buffer = []
|
||||
if stripped == "---":
|
||||
blocks.append(RawBlock("separator", "---", None, section_path, current_section_title))
|
||||
elif not stripped and not line.startswith('>'):
|
||||
if buffer:
|
||||
content = "\n".join(buffer).strip()
|
||||
if content: blocks.append(RawBlock("paragraph", content, None, section_path, current_section_title))
|
||||
buffer = []
|
||||
else:
|
||||
buffer.append(line)
|
||||
|
||||
|
|
|
|||
|
|
@ -1,135 +1,143 @@
|
|||
"""
|
||||
FILE: app/core/chunking/chunking_strategies.py
|
||||
DESCRIPTION: Strategie für atomares Sektions-Chunking v3.7.0.
|
||||
Garantiert Sektions-Integrität durch ein flexibles Toleranz-Limit.
|
||||
Kein Splitting von Sektionen, solange sie 'ungefähr' passen.
|
||||
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.
|
||||
"""
|
||||
import math
|
||||
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
|
||||
|
||||
# Toleranz-Faktor: Erlaubt Chunks, bis zu 15% über 'max' zu wachsen,
|
||||
# um eine Sektion vollständig zu erhalten.
|
||||
FLEX_FACTOR = 1.15
|
||||
|
||||
def _safe_estimate(text: str) -> int:
|
||||
"""Sicherere Token-Schätzung für MD/Deutsch (Faktor 3.0 statt 4.0)."""
|
||||
return max(1, math.ceil(len(text.strip()) / 3.0))
|
||||
|
||||
def _create_context_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
|
||||
parts = []
|
||||
if doc_title: parts.append(doc_title)
|
||||
def _create_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
|
||||
parts = [doc_title] if doc_title else []
|
||||
if sec_title and sec_title != doc_title: parts.append(sec_title)
|
||||
prefix = " > ".join(parts); return f"{prefix}\n{text}".strip() if prefix else text
|
||||
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, 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)
|
||||
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
|
||||
|
||||
# Das flexible Maximum, das Sektionen unzertrennt lässt
|
||||
soft_max = int(max_tokens * FLEX_FACTOR)
|
||||
|
||||
chunks: List[Chunk] = []
|
||||
|
||||
def _emit_chunk(block_list: List[RawBlock]):
|
||||
"""Schreibt eine Liste von Blöcken als einen einzigen Chunk ohne internes Splitting."""
|
||||
if not block_list: return
|
||||
txt = "\n\n".join([b.text for b in block_list])
|
||||
def _emit(txt, title, path):
|
||||
idx = len(chunks)
|
||||
title = block_list[0].section_title
|
||||
path = block_list[0].section_path
|
||||
win = _create_context_win(doc_title, title, txt)
|
||||
win = _create_win(doc_title, title, 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(txt),
|
||||
section_title=title, section_path=path,
|
||||
neighbors_prev=None, neighbors_next=None
|
||||
text=txt, window=win, token_count=estimate_tokens(txt),
|
||||
section_title=title, section_path=path, neighbors_prev=None, neighbors_next=None
|
||||
))
|
||||
|
||||
def _split_giant_section(sec_blocks: List[RawBlock]):
|
||||
"""Notfall-Split: Nur wenn eine EINZELNE Sektion bereits > soft_max ist."""
|
||||
full_text = "\n\n".join([b.text for b in sec_blocks])
|
||||
main_title = sec_blocks[0].section_title; main_path = sec_blocks[0].section_path
|
||||
header_text = sec_blocks[0].text if sec_blocks[0].kind == "heading" else ""
|
||||
# --- FALL A: HARD SPLIT (enable_smart_edge_allocation: false) ---
|
||||
if not smart_edge:
|
||||
buf = []
|
||||
for b in blocks:
|
||||
# Trenne bei jeder 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"):
|
||||
_emit("\n\n".join([x.text for x in buf]), buf[0].section_title, buf[0].section_path)
|
||||
buf = []
|
||||
buf.append(b)
|
||||
if buf: _emit("\n\n".join([x.text for x in buf]), buf[0].section_title, buf[0].section_path)
|
||||
return chunks
|
||||
|
||||
sents = split_sentences(full_text)
|
||||
cur_sents = []; sub_len = 0
|
||||
for s in sents:
|
||||
slen = _safe_estimate(s)
|
||||
if sub_len + slen > target and cur_sents:
|
||||
_emit_chunk([RawBlock("paragraph", " ".join(cur_sents), None, main_path, main_title)])
|
||||
ov_s = [header_text] if header_text else []
|
||||
ov_l = _safe_estimate(header_text) if header_text else 0
|
||||
for os in reversed(cur_sents):
|
||||
if os == header_text: continue
|
||||
t_len = _safe_estimate(os)
|
||||
if ov_l + t_len < overlap:
|
||||
ov_s.insert(len(ov_s)-1 if header_text else 0, os); ov_l += t_len
|
||||
else: break
|
||||
cur_sents = list(ov_s); cur_sents.append(s); sub_len = ov_l + slen
|
||||
else: cur_sents.append(s); sub_len += slen
|
||||
if cur_sents: _emit_chunk([RawBlock("paragraph", " ".join(cur_sents), None, main_path, main_title)])
|
||||
|
||||
# 1. Gruppierung in atomare Sektions-Einheiten
|
||||
# --- FALL B: SMART EDGE ALLOCATION (Pack-and-Carry-Over) ---
|
||||
# 1. Gruppierung in atomare Sektions-Einheiten (Sektions-Isolation)
|
||||
sections: List[List[RawBlock]] = []
|
||||
curr_sec: List[RawBlock] = []
|
||||
curr = []
|
||||
for b in blocks:
|
||||
if b.kind == "heading" and b.level <= split_level:
|
||||
if curr_sec: sections.append(curr_sec)
|
||||
curr_sec = [b]
|
||||
else: curr_sec.append(b)
|
||||
if curr_sec: sections.append(curr_sec)
|
||||
if curr: sections.append(curr)
|
||||
curr = [b]
|
||||
else: curr.append(b)
|
||||
if curr: sections.append(curr)
|
||||
|
||||
# 2. Das flexible Pack-Verfahren
|
||||
current_chunk_buf: List[RawBlock] = []
|
||||
current_tokens = 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]
|
||||
current_chunk_text = ""
|
||||
current_meta = {"title": None, "path": "/"}
|
||||
|
||||
for sec in sections:
|
||||
sec_text = "\n\n".join([b.text for b in sec])
|
||||
sec_tokens = _safe_estimate(sec_text)
|
||||
while processing_queue:
|
||||
item = processing_queue.pop(0)
|
||||
item_text = item["text"]
|
||||
item_tokens = estimate_tokens(item_text)
|
||||
|
||||
if current_chunk_buf:
|
||||
# PRÜFUNG: Würde die neue Sektion das FLEXIBLE Limit sprengen?
|
||||
if (current_tokens + sec_tokens > soft_max):
|
||||
_emit_chunk(current_chunk_buf)
|
||||
current_chunk_buf = []
|
||||
current_tokens = 0
|
||||
# Haben wir das Ziel-Maß erreicht und es kommt eine neue Sektion?
|
||||
elif (current_tokens >= target):
|
||||
_emit_chunk(current_chunk_buf)
|
||||
current_chunk_buf = []
|
||||
current_tokens = 0
|
||||
# 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
|
||||
|
||||
# Wenn eine EINZELNE Sektion alleine schon das weiche Limit sprengt
|
||||
if not current_chunk_buf and sec_tokens > soft_max:
|
||||
_split_giant_section(sec)
|
||||
combined_est = estimate_tokens(current_chunk_text + "\n\n" + item_text) if current_chunk_text else item_tokens
|
||||
|
||||
# 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()
|
||||
else:
|
||||
current_chunk_buf.extend(sec)
|
||||
current_tokens = _safe_estimate("\n\n".join([b.text for b in current_chunk_buf]))
|
||||
# Regel 2: Wenn Puffer voll -> Emittieren und Sektion zurücklegen
|
||||
if current_chunk_text:
|
||||
_emit(current_chunk_text, current_meta["title"], current_meta["path"])
|
||||
current_chunk_text = ""
|
||||
processing_queue.insert(0, item)
|
||||
else:
|
||||
# Regel 3: Einzelne Sektion zu groß -> 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
|
||||
|
||||
if current_chunk_buf:
|
||||
_emit_chunk(current_chunk_buf)
|
||||
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"])
|
||||
|
||||
# Carry-Over: Der Rest wird an den Anfang der Queue geschoben
|
||||
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})
|
||||
|
||||
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."""
|
||||
target = config.get("target", 400); max_tokens = config.get("max", 600)
|
||||
chunks: List[Chunk] = []; buf: List[RawBlock] = []
|
||||
|
||||
for b in blocks:
|
||||
b_tokens = _safe_estimate(b.text)
|
||||
current_tokens = sum(_safe_estimate(x.text) for x in buf) if buf else 0
|
||||
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:
|
||||
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))
|
||||
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
|
||||
chunks.append(Chunk(id=f"{note_id}#c{idx:02d}", note_id=note_id, index=idx, text=txt, window=win, token_count=_safe_estimate(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
|
||||
Loading…
Reference in New Issue
Block a user