Concept Contrast Fable

Reranking and avoiding results that are too similar

Conclusion first:Reranker reorders candidate retrieval results by relevance to improve quality, while MMR balances relevance and diversity to avoid duplicate results. The two are often conflated in the post-retrieval stage, but they solve different problems.

Reranker reorders candidate retrieval results by relevance to improve quality, while MMR balances relevance and diversity to avoid duplicate results. The two are often conflated in the post-retrieval stage, but they solve different problems.

Fable

Two shopkeepers, one lining up the closest match, one deliberately diverging

On East Street in the county town there was a general store run by a shopkeeper named Zheng. The shop had many supply channels, and whatever customers asked, Zheng could pull a dozen items from the storeroom. Later customers began complaining: the first ten things you find are almost identical—we can't tell the difference. Zheng thought it over and hired two helpers to do the filtering.

Helper A was in charge of ranking. He calculated the similarity between each item and the customer's needs, then queued them from highest to lowest score. When the customer asked about cloth shoes, he found twenty items, and the ones at the front were the cheapest, the cheapest, the cheapest—because price was closest to the customer's mental price point, and all other attributes were overshadowed by price.

Helper B was responsible for diversity. He specifically split similar things apart, reasoning that if customers did not see more options, they would not know which one truly suited them. Once a customer asked about cloth shoes, and he deliberately placed an expensive pair next to a cheap pair, saying this gave the customer room for comparison. The customer found it baffling—the expensive one was clearly better, so why put it with the cheap one.

Manager Zheng observed for a month and found that Helper A and Helper B each had problems when used alone. With Helper A's ranking, customers always said the recommendations were repetitive and showed no difference. With Helper B's diversity, customers always asked why worse items were placed first when there were clearly better ones.

Manager Zheng called the two of them together to study it. The conclusion was: first use Helper B's method to pick out several differentiated directions from the candidates, then within each direction use Helper A's similarity ranking. This ensured both diversity and that what was shown within each direction was the closest to the need.

Helper A had an objection to this plan: then what I do hasn't changed, it's still ranking, just ranked after diversity. Diversity is Helper B's contribution. Manager Zheng said the two of you are doing different things: ranking orders by similarity, while diversity manages whether there is value in other dimensions beyond similarity. What the guest wants is not just sameness, but the most right choice among the options worth seeing.

Later this method became the shop's standard process: first pass the diversity filter, then do relevance ranking. Complaints about duplicate questions decreased, and Helper A and Helper B stopped complaining about each other—because each person's work had independent value; they just needed a reasonable order to string them together.

Reranker

A model that reranks candidate retrieval results by relevance, usually computing relevance scores based on richer semantic information. It can improve coarse recall results, but does not consider result diversity.

MMR maximum marginal relevance

Introduce diversity constraints beyond relevance, deliberately introducing differences among candidate results to avoid duplication among the most relevant results. Suitable for scenarios that need to show multiple dimensions, but may rank slightly less relevant results higher.

Story Mapping

Helper A sorts by similarity: Reranker, which reorders candidate results by relevance
Helper B deliberately separates similar goods: MMR, introducing a diversity constraint on top of relevance
Ranking only the top ten by similarity yields nearly identical items: the Reranker does not handle duplication, and information richness does not improve
Seeking only diversity puts the bad ones first: MMR leans too far toward diversity, and relevance suffers
First pass diversity, then rank within each direction: MMR + Reranker combination, filter first and optimize later
Sorting governs similarity, diversity governs whether other dimensions exist: the two mechanisms solve different problems, and only when combined in order can both be addressed

Post-processing design once it lands in the retrieval system

When optimizing retrieval results, first confirm what your problem is. If users complain they cannot find the content they want or that ranking is unreasonable, use a Reranker—it can put more relevant content first. If users complain that results are repetitive and indistinguishable, use MMR—it forces diversity into the candidates. The most effective approach is to combine both: first use MMR for diversity filtering, then use a Reranker to rank relevance within each direction. This order cannot be reversed—ranking first and then diversifying will destroy the Reranker's optimization results. Diversifying first and then ranking ensures both diversity and quality within each direction.

Reranker Maximum Marginal Relevance