Concept Contrast Fable

Reranking and reranker: one describes what was done, the other describes what was used to do it

Conclusion first:When doing retrieval augmentation, people often discuss reranking and rerankers as if they were the same thing. Reranking is a processing step in the search pipeline that reorders initial results by relevance; a reranker is the specific model or algorithm that performs this step. When discussing technical solutions, first distinguish: are you missing this step, or are you missing a better ranking model?

When doing retrieval augmentation, people often discuss reranking and rerankers as if they were the same thing. Reranking is a processing step in the search pipeline that reorders initial results by relevance; a reranker is the specific model or algorithm that performs this step. When discussing technical solutions, first distinguish: are you missing this step, or are you missing a better ranking model?

Fable

The pharmacy's medicine-dispensing process

A traditional Chinese medicine pharmacy handled several hundred prescriptions a day. The old pharmacist had a habit: before filling a prescription, he would lay out all the herbs on the counter and arrange them by the importance of their dosage in the formula—herbs with large doses and primary effects placed within easiest reach, auxiliary ingredients and guides placed in the back row. Only after arranging them would he begin bagging.

The pharmacy installed a new automatic sorting machine, and the manufacturer said it could automatically arrange medicines by the first letter of their pinyin. The pharmacist tried it for two days and found something wrong—the machine put Dangshen and Danggui together because both start with "Dang," but in a prescription Dangshen is the main ingredient while Danggui is only an auxiliary. Filling a prescription in the machine's order, the main ingredient was hidden on the third row while the auxiliary was at the very front, making bagging a scramble.

The pharmacist argued with the supervisor for a week. The supervisor said the sorting machine was already very advanced—what more do you want? The pharmacist said the sorting machine could indeed arrange medicines, but the basis for arranging them was wrong. What I want is not a machine that can arrange medicines, but a machine that can arrange them correctly by prescription importance. Only then did the supervisor understand: the step of arranging medicines itself was fine; it was the thing executing the arrangement that was wrongly chosen.

The pharmacy finally set two rules. First, the medicine-arranging step must be retained—loading directly from the sorting machine into bags saved the arranging step, but the wrong-bag rate tripled. Second, the person or machine arranging the medicine must use the prescription dosage as the sorting basis, and may not use pinyin, stroke count, or other rules unrelated to efficacy.

A year later the pharmacy upgraded its sorting system and swapped in a new ordering module. The process didn't change—still sort first, then fill—but the sorter went from an old pharmacist to an algorithm with dosage weights. Bagging efficiency rose forty percent, and the mis-bag rate didn't rise. In the procurement report, the manager specifically wrote: this upgrade replaced the sorter, it didn't add a new step.

Later, an unwritten procurement rule emerged in the pharmacy industry: when buying a sorting machine, do not write the medication-arranging function as one integrated feature. In the contract, split it into two lines: whether a medication-arranging step is needed, and the sorting algorithm of the medication-arranging module. Each line is priced, tested and accepted separately.

When the old pharmacist retired, he said one thing: first figure out whether you need someone to help you arrange the medicine, then figure out what kind of eyes that person uses to look at the medicine. If you lump whether to arrange and what to arrange with into one discussion, you will end up buying a very expensive sorting machine, and the medicine will still come out messy.

Reranking

A processing step in a retrieval pipeline that reorders preliminary retrieval results by relevance, importance, or other criteria; it is one stage in the process.

Reranker

The specific model, algorithm, or component that executes the reranking step, determining the logic, features, and weights on which the ranking is based, is the concrete implementation of this stage.

Story Mapping

The old apothecary sorts the medicines before filling a prescription: reranking, a processing step that is always present in the pipeline
An automatic sorting machine arranges medicines by the first letter of their pinyin: one implementation of a reranker, where the sorting criteria determine the quality of the arrangement
Keep the reranking step but change the ordering basis: keep the reranking step, upgrade the reranker model
The supervisor cannot tell whether to arrange medicine or change the sorting method: confusing step requirements with implementation selection during solution discussion

Applying it to an Agent project

When selecting a RAG system, write reranking as an independent pipeline node, clearly defining its input (the initial retrieval result list) and output (the reranked list). Then choose a specific reranker for this node. During evaluation, measure separately: the change in effectiveness after removing this node tells you whether the step has value, and the change in effectiveness after swapping in a different reranker tells you whether the model choice is right.

Reranking Reranker