← Back to Concept Contrast

Concept Contrast Fable

The big family of methods that split long text into chunks, and the one among them that slides along at a fixed step size

Conclusion first:Chunking is a broad category — fixed-length splitting, paragraph splitting, semantic-boundary splitting, sliding-window splitting; as long as long text is cut into several small chunks, it counts as chunking. Sliding-window chunking is a specific sub-strategy: fixed window length, fixed step size, with overlap between adjacent chunks. Treating sliding windows as all chunking methods loses the benefits of semantic-boundary splitting; treating paragraph splitting as sliding windows misses transition information at the boundaries. Which cutting method to choose depends on the material and the purpose.

Chunking is a broad category — fixed-length splitting, paragraph splitting, semantic-boundary splitting, sliding-window splitting; as long as long text is cut into several small chunks, it counts as chunking. Sliding-window chunking is a specific sub-strategy: fixed window length, fixed step size, with overlap between adjacent chunks. Treating sliding windows as all chunking methods loses the benefits of semantic-boundary splitting; treating paragraph splitting as sliding windows misses transition information at the boundaries. Which cutting method to choose depends on the material and the purpose.

Fable

More than one knife in the rice cake shop: push-cutting is just one technique

Old Zhou of the rice cake shop was recognized in town for his excellent knife skills. He had three ways of cutting rice cakes. Straight cut — cut one rice cake into equal-width slices all the way through. Around-the-filling cut — cut along the positions of the red dates and sweet beans inside the rice cake, so each piece is uneven in size but contains one complete filling. Push-window cut — first cut a one-inch-long piece, then move back half a centimeter and cut the next piece, so the pieces overlap.

A restaurant ordered rice cakes for soup base, requiring two pieces per bowl, each piece able to hold minced meat and mushrooms from the soup. Old Zhou first tried the wrap-filling cut—each piece enclosed a red date, with the best appearance, but the dates all fell apart when boiled in soup, leaving only glutinous rice skin when scooped out. The restaurant said the taste was wrong. He switched to the push-window cut—each piece had overlapping surfaces, and the minced meat and mushrooms clung to the overlaps, making the soup taste rich. But when plated, every piece was the same length, making the presentation monotonous.

Old Zhou slowly assigned cutting methods by order. Sweet rice cake — soaked in sugar water, needing color and filling intact, cut by wrapping around the filling. Savory soup rice cake — with base ingredients clinging in chunks, needing overlapping surfaces to catch the soup, cut by pushing open a window. Rice cake platter — needing to be neat and attractive, stacked without collapsing, cut with a straight knife. All three cutting methods are cutting rice cake, but their fates after going into the pot are completely different.

Once the apprentice treated push-window cutting as the only way to cut. When he encountered a jujube rice cake, a one-chi-long cake with only three jujubes, the apprentice cut it by push-window cutting. Two jujubes fell in the overlap area, and after coming out of the pot the jujube flesh had been repeatedly heated and fell apart; the other was cut into two halves, and the residue mixed into the soup base. Old Zhou had the apprentice recut that cake by wrap-filling cutting—cutting one piece along each of the three jujubes, pieces large and small but the jujubes intact.

Another mismatch happened with the glutinous rice cake—a cake steamed from a mix of glutinous and non-glutinous rice, with a loose texture. The apprentice cut it by pushing the knife through, and adjacent pieces overlapped too much on the cut face; after going into the pot, they all stuck together into one lump. Old Zhou said this kind of cake had to be cut with a straight downward cut, one clean cut without going back.

The restaurant chef later explained the use before ordering. Old Zhou quoted the cutting method based on use—sweet or salty, plated or coated, firm cake or loose. The push-window cut became the brightest knife in Old Zhou's toolbox, but not the only one.

At year's end, Old Zhou summed it up for the apprentice: chunking is a door handle, and push-window cutting is one of the sharp knives within it. For hard pastries use the knife; for soft ones use your hands. With filling, go around; without filling, cut straight down. Anyone who walks in asking how to use push-window cutting on every rice cake gets told to boil a hundred catties of rice first.

Chunking

A broad category of general strategies for splitting long text into several short text chunks. It includes multiple sub-methods such as fixed-length splitting, paragraph-based splitting, semantic-boundary splitting, and sliding-window splitting, and the choice depends on the downstream task's specific requirements for chunk boundaries and chunk integrity.

Sliding Window Chunking

A specific chunking strategy—fixed window length and stride, with overlapping regions between adjacent chunks. The advantage is strong continuity of information between chunks, suited to scenarios that need to preserve contextual flow, but chunks may cut through semantic units, and the overlap also brings extra storage and processing overhead.

Story Mapping

  • Old Zhou's rice cake shop: a unified entry point for chunking strategies, with many ways to cut
  • Straight-cut slicing: fixed-length chunking that ignores content and semantic boundaries
  • Cutting around the filling: chunking by semantic boundaries to preserve the integrity of content units
  • Sliding window chunking: sliding-window chunking with fixed window and stride, with overlap between chunks
  • The apprentice cuts jujube cakes with a push-window: the classic mistake of treating the sliding window as the only method
  • The rice-pulp cake stuck completely when cut through the window: the material texture was unsuitable for the overlap strategy, and the cutting method needed to change

Applied to RAG system chunking strategy selection

For general documents, fixed-length chunking plus paragraph boundary correction is enough. Legal and contract texts that need to preserve paragraph semantic integrity should be chunked by semantic boundaries. Technical documents that require reasoning across chunk boundaries should use sliding-window chunking to ensure context continuity. Do not default to sliding windows right away—first understand the document structure and downstream retrieval needs, then decide window length and overlap rate. A high overlap rate does reduce sentence-break errors, but every extra inch of overlap consumes more storage and inference budget, and this gap widens on a million-scale document library.

京ICP备2026014272号