WP13b Refactoring ingestion und Chunker
This commit is contained in:
1 parent
cf302e8334
commit
94e5ebf577
13 files changed
+607
-762
No files matched your search
@@ -0,0 +1,74 @@
|
||||
"""
|
||||
FILE: app/core/chunking/chunking_strategies.py
|
||||
DESCRIPTION: Implementierung der mathematischen Splitting-Strategien.
|
||||
"""
|
||||
from typing import List, Dict, Any
|
||||
from .chunking_models import RawBlock, Chunk
|
||||
from .chunking_utils import estimate_tokens
|
||||
from .chunking_parser import split_sentences
|
||||
|
||||
def strategy_sliding_window(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, context_prefix: str = "") -> List[Chunk]:
|
||||
"""Fasst Blöcke zusammen und schneidet bei 'target' Tokens."""
|
||||
target = config.get("target", 400); max_tokens = config.get("max", 600)
|
||||
overlap_val = config.get("overlap", (50, 80))
|
||||
overlap = sum(overlap_val) // 2 if isinstance(overlap_val, tuple) else overlap_val
|
||||
chunks = []; buf = []
|
||||
|
||||
def _add(txt, sec, path):
|
||||
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=sec, section_path=path, neighbors_prev=None, neighbors_next=None))
|
||||
|
||||
def flush():
|
||||
nonlocal buf
|
||||
if not buf: return
|
||||
text_body = "\n\n".join([b.text for b in buf])
|
||||
sec_title = buf[-1].section_title; sec_path = buf[-1].section_path
|
||||
if estimate_tokens(text_body) <= max_tokens: _add(text_body, sec_title, sec_path)
|
||||
else:
|
||||
sents = split_sentences(text_body); cur_sents = []; cur_len = 0
|
||||
for s in sents:
|
||||
slen = estimate_tokens(s)
|
||||
if cur_len + slen > target and cur_sents:
|
||||
_add(" ".join(cur_sents), sec_title, sec_path)
|
||||
ov_s = []; ov_l = 0
|
||||
for os in reversed(cur_sents):
|
||||
if ov_l + estimate_tokens(os) < overlap: ov_s.insert(0, os); ov_l += estimate_tokens(os)
|
||||
else: break
|
||||
cur_sents = list(ov_s); cur_sents.append(s); cur_len = ov_l + slen
|
||||
else: cur_sents.append(s); cur_len += slen
|
||||
if cur_sents: _add(" ".join(cur_sents), sec_title, sec_path)
|
||||
buf = []
|
||||
|
||||
for b in blocks:
|
||||
if b.kind == "heading": continue
|
||||
if estimate_tokens("\n\n".join([x.text for x in buf])) + estimate_tokens(b.text) >= target: flush()
|
||||
buf.append(b)
|
||||
if estimate_tokens(b.text) >= target: flush()
|
||||
flush()
|
||||
return chunks
|
||||
|
||||
def strategy_by_heading(blocks: List[RawBlock], config: Dict[str, Any], note_id: str, doc_title: str = "") -> List[Chunk]:
|
||||
"""Splittet Text basierend auf Markdown-Überschriften."""
|
||||
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)
|
||||
chunks = []; buf = []; cur_tokens = 0
|
||||
|
||||
def _flush(title, path):
|
||||
nonlocal buf, cur_tokens
|
||||
if not buf: return
|
||||
txt = "\n\n".join(buf); win = f"# {doc_title}\n## {title}\n{txt}".strip() if title else txt
|
||||
idx = len(chunks)
|
||||
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))
|
||||
buf = []; cur_tokens = 0
|
||||
|
||||
for b in blocks:
|
||||
if b.kind == "heading":
|
||||
if b.level < split_level: _flush(b.section_title, b.section_path)
|
||||
elif b.level == split_level:
|
||||
if strict or cur_tokens >= target: _flush(b.section_title, b.section_path)
|
||||
continue
|
||||
bt = estimate_tokens(b.text)
|
||||
if cur_tokens + bt > max_tokens and buf: _flush(b.section_title, b.section_path)
|
||||
buf.append(b.text); cur_tokens += bt
|
||||
if buf: _flush(blocks[-1].section_title if blocks else None, blocks[-1].section_path if blocks else "/")
|
||||
return chunks
|
||||
Reference in new issue
Block a user