Chunking · Size Chunks Without Cutting Evidence

01 · 一条规则被切成两半 01 · One rule, cut in half

“标准版 30 天退货，但须未拆封”若刚好跨过切割线，检索到半句够用吗？ If “Standard: 30-day return, only when unopened” crosses a boundary, is half the sentence enough?

02 分块切割台：块长与重叠一起调 Chunk Cutter: Tune Length and Overlap Together

分块配方 chunk recipe

索引块数 Indexed chunks

块 chunks

证据完整度 Evidence completeness

100%

最佳块 best chunk

索引成本 Index cost

tokens

03 分块配方的四个后果 Four Consequences of a Chunk Recipe

粒度 · 完整 · 成本 granularity · context · cost

LENGTH

块长决定视野 Length sets the view

小块定位细，却容易割断一句完整规则。 Small chunks localize precisely but can sever a complete rule.

OVERLAP

重叠保护边界 Overlap protects boundaries

重复少量 token，让跨边界证据在相邻块里重现。 Repeat a few tokens so boundary-spanning evidence reappears in a neighboring chunk.

COUNT

步长决定块数 Stride sets chunk count

步长 = 块长 − 重叠；步长越小，块越多。 Stride equals length minus overlap; a smaller stride creates more chunks.

COST

重复增加成本 Duplication adds cost

重叠提高完整机会，也增加嵌入、存储和检索 token。 Overlap improves completeness odds while adding embedding, storage, and retrieval tokens.

04 最小机制：块长与步长共同排布边界 Minimal Mechanism: Length and Stride Place Boundaries

stride = L − O

n = ⌈(N − L) / (L − O)⌉ + 1

N 是文档 token 数，L 是块长，O 是重叠。主交互中块数与成本都由这三个量直接算出。 N is document length, L chunk length, and O overlap. The interaction computes chunk count and cost directly from these values.

选择 L 与 O Choose L and O 得到步长 derive stride

铺开切割线 Place boundaries 形成索引块 form index chunks

测完整与成本 Measure context and cost 不是只看块数 not chunk count alone

05 边界实验：小块且零重叠 Boundary Test: Tiny Chunks with Zero Overlap

主动制造失败 Create a failure

让切割线穿过关键规则 Cut through the decisive rule

块长 4、重叠 0 会把“30 天”和“未拆封”分到不同块。任一命中都只得到半条规则。 Length four with zero overlap separates “30 days” from “unopened.” Either hit retrieves only half the rule.

尚未执行。当前至少一个块包含完整规则。 Not run yet. At least one current chunk contains the full rule.

✗ 块越小，检索一定越准。 Smaller chunks always improve retrieval. → ✓ 粒度变细也可能割断必要上下文。 Finer granularity can sever required context.

✗ 重叠越大越保险。 More overlap is always safer. → ✓ 它也复制 token，增加索引与上下文成本。 It duplicates tokens and raises indexing and context cost.

✗ 固定字数适合所有文档。 One fixed length fits every document. → ✓ 结构与语义边界也应参与分块。 Structural and semantic boundaries should inform chunking.

06 · 一句话带走 06 · One line to keep

分块用块长与重叠在召回粒度、证据完整性和 token 成本之间取舍；边界切在哪里会决定检索能否拿到完整事实。 Chunking trades off recall granularity, evidence completeness, and token cost through length and overlap; boundary placement decides whether retrieval can recover a complete fact.