133 lines
6.2 KiB
Python
133 lines
6.2 KiB
Python
"""
|
|
FILE: app/core/chunking/chunking_strategies.py
|
|
DESCRIPTION: Universelle Strategie für atomares Sektions-Chunking v3.6.0.
|
|
Garantiert Sektions-Integrität durch präventives Chunk-Management.
|
|
"""
|
|
import math
|
|
from typing import List, Dict, Any, Optional
|
|
from .chunking_models import RawBlock, Chunk
|
|
from .chunking_parser import split_sentences
|
|
|
|
def _accurate_estimate_tokens(text: str) -> int:
|
|
"""Konservative Schätzung für deutschen Text (len/2.5 statt len/4)."""
|
|
return max(1, math.ceil(len(text.strip()) / 2.5))
|
|
|
|
def _create_context_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
|
|
parts = []
|
|
if doc_title: parts.append(doc_title)
|
|
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
|
|
|
|
def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, doc_title: str = "") -> List[Chunk]:
|
|
"""
|
|
Sektions-Chunking: Packt komplette Abschnitte in Chunks.
|
|
Bei Überlauf wird die Sektion ohne Ausnahme in den nächsten Chunk geschoben.
|
|
"""
|
|
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_chunk(block_list: List[RawBlock]):
|
|
"""Schreibt eine Liste von Blöcken als einen einzigen, ungeteilten Chunk."""
|
|
if not block_list: return
|
|
txt = "\n\n".join([b.text for b in block_list])
|
|
idx = len(chunks)
|
|
title = block_list[0].section_title
|
|
path = block_list[0].section_path
|
|
win = _create_context_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=_accurate_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 > 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 ""
|
|
|
|
sents = split_sentences(full_text)
|
|
cur_sents = []; sub_len = 0
|
|
|
|
for s in sents:
|
|
slen = _accurate_estimate_tokens(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 = _accurate_estimate_tokens(header_text) if header_text else 0
|
|
for os in reversed(cur_sents):
|
|
if os == header_text: continue
|
|
t_len = _accurate_estimate_tokens(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 Einheiten
|
|
sections: List[List[RawBlock]] = []
|
|
curr_sec: List[RawBlock] = []
|
|
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)
|
|
|
|
# 2. Das Pack-Verfahren (Kein Zerschneiden beim Flashen!)
|
|
candidate_chunk: List[RawBlock] = []
|
|
candidate_tokens = 0
|
|
|
|
for sec in sections:
|
|
sec_text = "\n\n".join([b.text for b in sec])
|
|
sec_tokens = _accurate_estimate_tokens(sec_text)
|
|
|
|
# Prüfung: Passt die Sektion noch dazu?
|
|
if candidate_tokens + sec_tokens <= max_tokens:
|
|
candidate_chunk.extend(sec)
|
|
candidate_tokens = _accurate_estimate_tokens("\n\n".join([b.text for b in candidate_chunk]))
|
|
else:
|
|
# Chunk ist voll -> Abschluss an Sektionsgrenze
|
|
if candidate_chunk:
|
|
_emit_chunk(candidate_chunk)
|
|
candidate_chunk = []
|
|
candidate_tokens = 0
|
|
|
|
# Neue Sektion allein prüfen
|
|
if sec_tokens > max_tokens:
|
|
_split_giant_section(sec)
|
|
else:
|
|
candidate_chunk = list(sec)
|
|
candidate_tokens = sec_tokens
|
|
|
|
if candidate_chunk: _emit_chunk(candidate_chunk)
|
|
return chunks
|
|
|
|
def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, context_prefix: str = "") -> List[Chunk]:
|
|
target = config.get("target", 400); max_tokens = config.get("max", 600)
|
|
chunks: List[Chunk] = []; buf: List[RawBlock] = []
|
|
for b in blocks:
|
|
b_tokens = _accurate_estimate_tokens(b.text)
|
|
current_tokens = sum(_accurate_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=_accurate_estimate_tokens(txt), section_title=buf[0].section_title, section_path=buf[0].section_path, neighbors_prev=None, neighbors_next=None))
|
|
return chunks |