How a Fruit Store Story Reveals the Secrets of Search Recall and Precision

Fruit Store Story

One day you see a goddess eating a white strawberry that looks delicious. You go to a fruit shop, ask the owner for strawberries, and he shows several varieties. The labels reveal that none of them are the white strawberry you want, but later you discover the exact type among the fruits the owner hadn't initially displayed.

The owner’s poor memory of high‑profit fruits leads to missed opportunities.

Recall and Precision in Search

The fruit shop scenario maps to a typical search situation. All fruit types can be divided into four categories:

A and B are the fruits the owner displayed; C and D are the fruits he did not. A and C are strawberry‑related, B and D are not.

Recall = retrieved relevant results / all relevant results = A / (A + C). Precision = retrieved relevant results / all retrieved results = A / (A + B).

The fruit shop owner’s recall is 75 % and precision is 60 %, which explains the owner’s frustration.

Recruitment Search Dilemma

Recruitment search faces the same problem: many relevant results stay in region C (relevant but not retrieved). Employers get poor paid‑search performance, job seekers see fewer opportunities, and the platform’s match rate and revenue drop.

To improve the situation we need to:

Move information from C to A (increase recall).

Prevent D‑region items from entering B (maintain precision).

Although recall and precision are mathematically independent, in practice loosening retrieval strategies to raise recall often harms precision, so both must be balanced.

Search Design Patterns

We compiled a matrix of design patterns that can improve recall and precision at different search stages.

The vertical axis shows the emphasis on recall (top) or precision (bottom); the horizontal axis shows search stages. Specific patterns can be selected according to needs.

Design Pattern – Cross‑Database / Cross‑Category Search

Cross‑database search means querying multiple databases or datasets simultaneously; cross‑category search means searching across different categories. This is crucial when users do not know where to start.

58’s overly fine‑grained category hierarchy forces users to click through multiple levels, reducing search efficiency.

We adopted three classification‑search approaches (search box, assist area, result‑page category selector) and improved the result‑page selector to reduce decision pressure while showing more results.

We expanded the searchable categories (Old showed only the matched secondary category; New shows all recruitment categories) and kept the category‑switch entry for consistency.

We also integrated third‑level categories into the second‑level switch to lower cognitive load.

To compensate for reduced precision after expanding categories, we added recommended tags at the top and in the feed, leveraging recommendation algorithms.

Design Pattern – Preset Search Sets

Machine‑generated results can have flaws; manual presets improve recall and precision. We pre‑match keyword‑result sets and place them in recommendation or classification entrances, continuously refining them with feedback.

For example, the keyword “delivery person” spans logistics, catering, retail, and admin categories. We create a virtual category containing results from all these areas and also surface related terms such as “courier”, “food delivery”, etc.

Design Pattern – Matching Labels

Matching labels highlight keyword‑related information on relevant results, improving explainability and perceived accuracy, helping users quickly lock onto relevant items.

We use semantic analysis and intent recognition to personalize job cards: titles, tags, and recommendation reasons are dynamically generated based on user intent and job attributes.

Design Pattern – Tag Collection

Many of these patterns rely on recommendation technology. Using accurate recommendation tags enhances recall precision, and building a comprehensive user‑tag system is essential.

Other Applications of Search Design Patterns

Additional patterns applied in the redesign include auto‑suggest, collection/subscription, pagination, sorting optimization, and search‑result feedback.

Design Review

The core goal of the search‑list redesign was to improve connection efficiency and conversion rate while keeping precision.

We decomposed the business goal: increase recall without sacrificing precision. Strategies included cross‑database/category search, virtual categories, preset search sets, and UI optimizations.

We also enhanced card styles and matching labels, using intelligent title stitching, tags, and recommendation reasons, and collected user‑tag data for feedback loops.

Six gray‑scale tests showed that each factor (breaking category limits, title stitching, style changes) improved recall while maintaining precision, with the best variant doubling key metrics.

Conclusion

The search design pattern matrix is a practical model for boosting recall and precision; each pattern offers distinct strategies that can be adapted to specific scenarios. Continuous iteration and feedback are essential for further refinement.