Concept Contrast Fable

One class of methods and one specific formula

Conclusion first:Sparse retrieval is an entire class of retrieval methods based on term frequency and document statistics; BM25 is the most commonly used specific probabilistic formula within it. Replacing the term sparse retrieval with BM25 in technology selection can lead to missing other sparse approaches such as TF-IDF and Boolean retrieval, or treating the limitations of a specific implementation as defects of the whole class during evaluation.

Sparse retrieval is an entire class of retrieval methods based on term frequency and document statistics; BM25 is the most commonly used specific probabilistic formula within it. Replacing the term sparse retrieval with BM25 in technology selection can lead to missing other sparse approaches such as TF-IDF and Boolean retrieval, or treating the limitations of a specific implementation as defects of the whole class during evaluation.

Fable

There is a saying in the carpenter guild in the south of the city: mortise-and-tenon joint. But Wang's shop uses dovetail joints

In the carpenter's guild in the south of the city, an ancestral craft called joinery has been passed down. Whether making tables, cabinets, or repairing roof beams, the master always tells apprentices: whenever wood is joined, it must be joined with joinery. To outsiders, joinery sounds like one thing. But those in the trade know there are more than a dozen types—dovetail joints, right-angle joints, through-tenon joints, miter joints—each for its own scenario.

Wang's carpentry shop used dovetail joints most of all—strong pull plates, drawers that wouldn't warp for two hundred years. Through three generations, Wang's reputation spread across half the province. Out-of-town merchants came to order furniture, specifically requesting joinery craftsmanship, and Wang's delivered with dovetail joints. Once, twice, a dozen times—no problems. The merchants eventually equated joinery with dovetail joints.

The problem arose with an order from a southern merchant. He wanted a set of movable stage platforms with mortise-and-tenon joints—assembled for each performance, dismantled afterward, loaded onto carts and shipped to the next town. Wang's shop built an entire framework using dovetail joints. When it arrived at the second town, all four corners had split. The dovetail joints fit too tightly; one disassembly damaged the wood, three disassemblies rendered it useless. The customer slammed the table.

Master Wang's senior craftsman was invited to look at the site. He said nothing, removed a crossbeam, and replaced it with a through-tenon. The through-tenon does not need to be locked tight; it is laterally secured by a wedge, and when it is time to dismantle, pulling out the wedge loosens it. The stage traveled through eight towns without falling apart. The client asked what the difference was. The senior craftsman said: dovetail joints emphasize sturdiness, through-tenons emphasize flexibility, and both are mortise-and-tenon joints. It is you who treated Master Wang as the entirety of mortise-and-tenon work.

Later, the carpenters' guild added a rule when taking out-of-town orders: when a client says mortise-and-tenon, the carpenter must ask about the scenario—fixed furniture or movable structure, load-bearing or panel, one-time installation or repeated disassembly. The dovetail joint is just one formula within mortise-and-tenon; used correctly it's unbeatable, used incorrectly it breaks faster than other joints.

The young carpenters in the shop made the same mistake when learning the craft. They saw that Wang's dovetail joint was the most famous, so they assumed joinery meant dovetail joints. The master had them look through old pattern books, and only then did they discover that before dovetail joints, right-angle joints, through-batten joints, and mitered joints had long had their own uses. If joinery is treated as dovetail joints, many old methods for specific scenarios are ignored.

The old rule of the carpenters' guild was carved on the door frame: mortise-and-tenon is the general term, dovetail joint is the specific term. You can't say mortise-and-tenon doesn't work, because that only accounts for one piece of wood.

Sparse retrieval

A class of retrieval methods based on term frequency and document statistics that represent queries and documents as sparse vectors. It includes many specific algorithms such as TF-IDF, BM25 and Boolean retrieval, and is not tied to any single formula.

BM25

The most commonly used probabilistic ranking function in sparse retrieval, based on term frequency saturation and document length normalization. It is a specific formula rather than an entire class of methods, and cannot represent all sparse retrieval approaches.

Story Mapping

Mortise-and-tenon joint: sparse retrieval, an entire class of methods based on term frequency and document statistics
Dovetail Joint: BM25, a specific probabilistic ranking formula
An out-of-town merchant directly treats mortise-and-tenon as a dovetail joint: the common misconception in engineering of equating sparse retrieval with BM25
The stage collapses: BM25 is not optimal in some scenarios; the tenon joints represent other sparse solutions
Boolean methods have been in use for thirty years: besides BM25, there are also mature approaches such as Boolean retrieval and TF-IDF
Inscription on the door frame: a common name is a category, a proper name is an instance; the distinction must be made when selecting a type

Down to retrieval solution selection

When writing a system design document, if you write sparse retrieval, it means you need to further specify which concrete algorithm to use; if you write BM25, it means you have already chosen the algorithm. Writing a general term as a specific name limits future selection space, and writing a specific name as a general term invites reviewers to press for implementation details. In real engineering, prioritize evaluating how well in-domain data fits each sparse approach—BM25's IDF assumption does not hold on all corpora.

Sparse retrieval BM25