Spaces:
Sleeping
Sleeping
File size: 9,361 Bytes
356cfcc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 | """层次化分块器.
策略:
1. 优先按 markdown 标题拆分 (H1/H2/H3)
2. 每个 section 内:
- 段间按 token 预算合并 (parent: 1500-2000t)
- 段内进一步切分 (child: 400-512t)
3. 可选: 语义分块 (按 sentence 相似度二次切)
4. 可选: Anthropic-style 上下文预置 (LLM 给每个 chunk 加 context prefix)
输出:
- parents: 大块, 给 LLM 喂上下文用
- children: 小块, 给向量检索用
- 每个 child 带 parent_id, page_no, heading 字段
"""
from __future__ import annotations
import asyncio
import logging
import re
import time
import uuid
from dataclasses import dataclass, field
from app.config import settings
from app.llm.base import LLMMessage
from app.services.parsers.base_parser import ParsedDocument
logger = logging.getLogger(__name__)
# 标题正则 (markdown)
_HEADING_RE = re.compile(r"^(#{1,6})\s+(.+?)\s*$", re.MULTILINE)
@dataclass
class Chunk:
"""分块结果 (child or parent)."""
id: str
doc_id: str
parent_id: str | None
chunk_index: int
text: str
token_count: int
page_no: int | None = None
heading: str | None = None
context_prefix: str | None = None
created_at: float = field(default_factory=time.time)
def _approx_token_count(s: str) -> int:
"""粗估 token 数. 用 chars/2 当近似, 避免引 tiktoken."""
if not s:
return 0
# 中英文混合: CJK 算 1.5 token/char, 其它算 0.5
cjk = sum(1 for c in s if "一" <= c <= "鿿")
other = len(s) - cjk
return int(cjk * 1.5 + other * 0.5)
def _split_by_headings(markdown: str) -> list[tuple[str, str]]:
"""按 markdown 标题拆分. 返回 [(heading, body), ...].
heading 为 None 表示标题前的引言段.
"""
matches = list(_HEADING_RE.finditer(markdown))
if not matches:
return [(None, markdown)]
chunks: list[tuple[str, str]] = []
# 引言段
pre = markdown[: matches[0].start()].strip()
if pre:
chunks.append((None, pre))
for i, m in enumerate(matches):
heading = m.group(2).strip()
start = m.end()
end = matches[i + 1].start() if i + 1 < len(matches) else len(markdown)
body = markdown[start:end].strip()
if body:
chunks.append((heading, body))
return chunks
def _sliding_window(text: str, target: int, overlap: int) -> list[str]:
"""按 token 估值的滑动窗口切分.
简化版: 按段落切, 累积到 target token 就 flush, 与上一段重叠 overlap.
"""
paragraphs = [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
if not paragraphs:
return []
pieces: list[str] = []
buf: list[str] = []
buf_tokens = 0
for p in paragraphs:
pt = _approx_token_count(p)
# 单段超过 target, 硬切
if pt > target:
if buf:
pieces.append("\n\n".join(buf))
buf, buf_tokens = [], 0
# 字符级切
step = max(int(target * 2), 200) # target * 2 chars ≈ target tokens
for i in range(0, len(p), step):
pieces.append(p[i : i + step])
continue
if buf_tokens + pt > target and buf:
pieces.append("\n\n".join(buf))
# overlap: 保留最后一段
if overlap > 0 and buf:
last = buf[-1]
last_t = _approx_token_count(last)
if last_t <= overlap:
buf = [last]
buf_tokens = last_t
else:
buf, buf_tokens = [], 0
else:
buf, buf_tokens = [], 0
buf.append(p)
buf_tokens += pt
if buf:
pieces.append("\n\n".join(buf))
return pieces
def _split_by_paragraphs(text: str) -> list[str]:
return [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
@dataclass
class ChunkingResult:
parents: list[Chunk]
children: list[Chunk]
def chunk_document(parsed: ParsedDocument, doc_id: str) -> ChunkingResult:
"""对 ParsedDocument 做层次化分块.
步骤:
1. 按 markdown 标题分 section
2. 每个 section 用滑动窗口切 parent (2000t, overlap 200)
3. 每个 parent 内用滑动窗口切 child (512t, overlap 64)
4. 按 page 编号 (按 parent 在 markdown 中的字符位置近似)
"""
md = parsed.markdown
sections = _split_by_headings(md)
page_text = parsed.pages # 可能为空 (Marker)
parents: list[Chunk] = []
children: list[Chunk] = []
chunk_idx = 0
for heading, body in sections:
# 先切 parent
for ptext in _sliding_window(body, target=2000, overlap=200):
pid = uuid.uuid4().hex
ptoks = _approx_token_count(ptext)
# 估算 page_no
page_no = _estimate_page_no(ptext, parsed) if page_text else None
parent = Chunk(
id=pid,
doc_id=doc_id,
parent_id=None,
chunk_index=chunk_idx,
text=ptext,
token_count=ptoks,
page_no=page_no,
heading=heading,
)
parents.append(parent)
chunk_idx += 1
# 切 child
for ctext in _sliding_window(ptext, target=settings.chunk_size, overlap=settings.chunk_overlap):
cid = uuid.uuid4().hex
ctoks = _approx_token_count(ctext)
child = Chunk(
id=cid,
doc_id=doc_id,
parent_id=pid,
chunk_index=chunk_idx,
text=ctext,
token_count=ctoks,
page_no=page_no,
heading=heading,
)
children.append(child)
chunk_idx += 1
return ChunkingResult(parents=parents, children=children)
def _estimate_page_no(snippet: str, parsed: ParsedDocument) -> int | None:
"""在 parsed.pages 里找包含 snippet 片段的页. 简单字符串包含."""
if not parsed.pages:
return None
# 取 snippet 前 50 字符作锚
anchor = snippet[:50].strip()
if not anchor:
return None
for p in parsed.pages:
if anchor in p.text:
return p.page_no
return None
# ========== 上下文预置 (Anthropic-style) ==========
_CONTEXT_PROMPT_TEMPLATE = (
"你是一名文档分块上下文标注助手。给定一段来自文档的 chunk (用 <chunk> 包裹) "
"以及文档标题, 请用 1-2 句中文描述该 chunk 在整篇文档中的上下文, 帮助后续检索时理解其含义。\n\n"
"要求:\n"
"- 简洁, 不超过 80 字\n"
"- 包含: 这是什么类型的内容 (定义 / 例子 / 数据 / 结论 / 步骤等), 在文档哪个章节\n"
"- 不要重复 chunk 原内容\n"
"- 仅输出描述本身, 不要加任何前缀\n\n"
"文档标题: {doc_title}\n"
"所属章节: {heading}\n"
"<chunk>\n{chunk}\n</chunk>\n\n"
"上下文描述:"
)
async def contextualize_chunks(
chunks: list[Chunk],
*,
doc_title: str = "未知文档",
max_concurrency: int = 4,
) -> list[Chunk]:
"""为每个 chunk 生成 context_prefix. Anthropic-style.
优化:
- 并发限流 (max_concurrency), 不爆 LLM 限流
- 失败容错: 单条失败不中断, 留 prefix=None
"""
if not settings.contextual_retrieval or not chunks:
return chunks
try:
from app.llm.factory import get_llm
except Exception as e: # noqa: BLE001
logger.warning("LLM not available for contextual retrieval: %s", e)
return chunks
sem = asyncio.Semaphore(max_concurrency)
async def _one(c: Chunk) -> None:
if c.context_prefix:
return
async with sem:
try:
llm = get_llm()
prompt = _CONTEXT_PROMPT_TEMPLATE.format(
doc_title=doc_title,
heading=c.heading or "无",
chunk=c.text[:1200], # 限长, 防 LLM token 爆
)
resp = await llm.chat(
messages=[
LLMMessage(role="system", content="你是上下文标注助手。"),
LLMMessage(role="user", content=prompt),
],
temperature=0.2,
max_tokens=160,
)
prefix = (resp.content or "").strip().strip('"').strip("'")
if len(prefix) > 200:
prefix = prefix[:200]
c.context_prefix = prefix or None
except Exception as e: # noqa: BLE001
logger.warning("Contextual retrieval failed for chunk %s: %s", c.id[:8], e)
await asyncio.gather(*[_one(c) for c in chunks])
n_ok = sum(1 for c in chunks if c.context_prefix)
logger.info("Contextualized %d/%d chunks", n_ok, len(chunks))
return chunks
def chunk_to_embed_text(c: Chunk) -> str:
"""向量化用的文本: context_prefix + heading + body."""
parts: list[str] = []
if c.context_prefix:
parts.append(f"[Context: {c.context_prefix}]")
if c.heading:
parts.append(f"[Section: {c.heading}]")
parts.append(c.text)
return "\n".join(parts)
|