153 lines
6.6 KiB
Python
153 lines
6.6 KiB
Python
"""
|
|
FILE: app/core/chunking/chunking_strategies.py
|
|
DESCRIPTION: Strategien für atomares Sektions-Chunking (WP-15b konform).
|
|
v3.3.5: Garantiert atomare Sektionsgrenzen durch präventiven Flush.
|
|
"""
|
|
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
|
|
|
|
def _create_context_win(doc_title: str, sec_title: Optional[str], text: str) -> str:
|
|
"""Baut den Breadcrumb-Kontext für das Embedding-Fenster."""
|
|
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]:
|
|
"""
|
|
Gruppiert Blöcke zu Sektionen und hält diese atomar zusammen.
|
|
Nutzt striktes Look-Ahead, um das Zerschneiden von Sektionsübergängen zu verhindern.
|
|
"""
|
|
strict = config.get("strict_heading_split", False)
|
|
target = config.get("target", 400)
|
|
max_tokens = config.get("max", 600)
|
|
split_level = config.get("split_level", 2)
|
|
overlap_config = config.get("overlap", (50, 80))
|
|
overlap = sum(overlap_config) // 2 if isinstance(overlap_config, (list, tuple)) else overlap_config
|
|
|
|
chunks: List[Chunk] = []
|
|
buf: List[RawBlock] = []
|
|
cur_tokens = 0
|
|
|
|
def _add_chunk(txt, title, path):
|
|
idx = len(chunks)
|
|
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=estimate_tokens(txt),
|
|
section_title=title, section_path=path,
|
|
neighbors_prev=None, neighbors_next=None
|
|
))
|
|
|
|
def _flush_buffer():
|
|
nonlocal buf, cur_tokens
|
|
if not buf: return
|
|
|
|
main_title = buf[0].section_title
|
|
main_path = buf[0].section_path
|
|
full_text = "\n\n".join([b.text for b in buf])
|
|
actual_tokens = estimate_tokens(full_text)
|
|
|
|
# Falls die gruppierten Sektionen in das Limit passen
|
|
if actual_tokens <= max_tokens:
|
|
_add_chunk(full_text, main_title, main_path)
|
|
else:
|
|
# Nur wenn eine Sektion ALLEINE zu groß ist, wird intern gesplittet
|
|
sents = split_sentences(full_text)
|
|
cur_sents = []; sub_len = 0
|
|
header_text = buf[0].text if buf[0].kind == "heading" else ""
|
|
|
|
for s in sents:
|
|
slen = estimate_tokens(s)
|
|
if sub_len + slen > target and cur_sents:
|
|
_add_chunk(" ".join(cur_sents), main_title, main_path)
|
|
|
|
# Overlap-Erzeugung und Header-Injektion
|
|
ov_s = [header_text] if header_text else []
|
|
ov_l = estimate_tokens(header_text) if header_text else 0
|
|
for os in reversed(cur_sents):
|
|
if os == header_text: continue
|
|
t_len = 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:
|
|
_add_chunk(" ".join(cur_sents), main_title, main_path)
|
|
|
|
buf = []; cur_tokens = 0
|
|
|
|
# SCHRITT 1: Gruppierung in atomare Sektions-Einheiten (Heading + Paragraphs)
|
|
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)
|
|
|
|
# SCHRITT 2: Verarbeitung der Sektionen mit strenger Vorausschau
|
|
for sec in sections:
|
|
sec_text = "\n\n".join([b.text for b in sec])
|
|
sec_tokens = estimate_tokens(sec_text)
|
|
|
|
if buf:
|
|
# PRÜFUNG 1: Passt die gesamte neue Sektion noch in den Chunk (bis max)?
|
|
# PRÜFUNG 2: Wenn wir über target sind, fangen wir auf jeden Fall neu an.
|
|
if (cur_tokens + sec_tokens > max_tokens) or (cur_tokens >= target):
|
|
_flush_buffer()
|
|
# PRÜFUNG 3: Wenn strict-mode aktiv ist und ein split_level erreicht wurde
|
|
elif strict and sec[0].kind == "heading" and sec[0].level == split_level:
|
|
_flush_buffer()
|
|
|
|
buf.extend(sec)
|
|
# Token-Zähler basierend auf dem tatsächlichen Puffer-Text aktualisieren
|
|
cur_tokens = estimate_tokens("\n\n".join([b.text for b in buf]))
|
|
|
|
# Falls eine einzelne Sektion (selbst nach flush) schon zu groß ist
|
|
if cur_tokens >= max_tokens:
|
|
_flush_buffer()
|
|
|
|
_flush_buffer()
|
|
return chunks
|
|
|
|
def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, context_prefix: str = "") -> List[Chunk]:
|
|
"""Standard-Sliding-Window für flache Texte."""
|
|
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 = estimate_tokens("\n\n".join([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 = []
|
|
current_tokens = 0
|
|
|
|
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=estimate_tokens(txt), section_title=buf[0].section_title, section_path=buf[0].section_path, neighbors_prev=None, neighbors_next=None))
|
|
|
|
return chunks |