Learn
Exploring Vector Database Use Cases

Enhancing App Functionality: Optimizing Search with Vector Databases

Feb 16, 20247 min read

Vector databases revolutionize app development by enhancing search functionalities with their ability to perform fast, accurate, and semantic searches.

By Cahyo Subroto

Read the entire series

Introduction

The search functionality within an app has a critical role in application performance and user satisfaction. In today's landscape, where information overload is common, great search features can significantly enhance user experience by providing quick, relevant, and accurate results. This could improve user engagement, and users are less likely to abandon the app due to their inability to find what they seek. Moreover, the search function can improve performance by optimizing the data retrieval process and reducing latency, thereby achieving speed and responsiveness for users.

Vector databases are revolutionizing application search capabilities by introducing a paradigm shift in data indexing, querying, and retrieval. Unlike traditional databases that rely solely on keyword matching, vector databases leverage advanced techniques and algorithms like nearest neighbor search and dimensionality reduction techniques to capture the semantic meaning of data. By searching on vectors, these databases enable context-aware semantic searches, which can understand the conceptual and contextual nuances behind user queries. This approach is valuable in applications like e-commerce recommendation engines, where users can search for products using complex conditions like "stylish summer dresses suitable for a beach wedding" and retrieve highly accurate and relevant results from massive catalogs. Additionally, in multimedia content platforms, vector databases can enable users to search for visually similar images or videos based on their semantic content rather than relying on manual tagging or keyword descriptions. With the ability to perform these intricate searches over large datasets with remarkable accuracy and speed – up to 1000 times faster than traditional databases in some cases – vector databases are poised to be a game-changer for search features within applications, unlocking new possibilities for enhanced user experiences and data-driven insights.

Basic of Vector Databases

A vector database stores vector embeddings, a mathematical representation of unstructured data that captures the object’s semantic meaning. At a close glance, a vector is an array of numerical values expressing a floating point's location along several dimensions and might look something like this:


Embeddings: [array([-3.09392996e-02, -1.80662833e-02,  1.34775648e-02,  2.77156215e-02,
       -4.86349640e-03, -3.12581174e-02, -3.55921760e-02,  5.76934684e-03,
        2.80773244e-03,  1.35783911e-01,  3.59678417e-02,  6.17732145e-02,
...
       -4.61330153e-02, -4.85207550e-02,  3.13997865e-02,  7.82178566e-02,
       -4.75336798e-02,  5.21207601e-02,  9.04406682e-02, -5.36676683e-02],
      dtype=float32)]
Dim: 384 (384,)

In vector search, you must transform your data into vector embedding before storing them in a vector database. User queries are converted into vector embeddings to perform similarity searches on your data collection. After that, we get a list of similar vectors. Because these vector embeddings contain the semantic meaning of an object, the clustered vectors are similar and likely relevant to each other.

Vector databases excel at managing vector operations and processing complex, high-dimensional data, making them well-suited for applications that require efficient similarity searches over large datasets. One key feature that enables vector databases to enhance search functionalities is their ability to leverage approximate nearest neighbor (ANN) search techniques in conjunction with specialized index structures.

Many vector databases support various index types to facilitate fast and accurate similarity searches, including GPU-accelerated indexes for leveraging hardware acceleration, in-memory indexes for low-latency querying, disk-based indexes for handling larger datasets that exceed memory capacity, and hybrid indexes that combine memory and disk storage. Some popular index implementations include FAISS (Facebook AI Similarity Search), HNSW (Hierarchical Navigable Small World graphs), and DiskANN.

Improving Search with Vector Databases

There are various search function optimization techniques that vector databases use to achieve faster query response and improved relevance of search results in vector databases. We will use Milvus or Zilliz (managed Milvus database) as examples.

Faster Query Responses

Finding exact and fully accurate search results could hinder performance since it would require a search operation for every piece of data available. Instead, vector databases like Milvus use approximate search algorithms (ANN) that optimize retrieval performance without sacrificing accuracy.

To support ANN, Milvus can utilize advanced indexing techniques, from vector to scalar index, dramatically improve search performance and speed up query response time. These indexing methods efficiently organize high-dimensional vector spaces to reduce the search area for a query, enabling fast retrieval of search queries. This means that even in a collection of millions, billions, or even trillions of vectors, users can expect results in milliseconds.

If accuracy is a requirement, you can also tune your Milvus database to prioritize accuracy (recall) over latency or QPS.

Improved Relevance of the Search Results

Vector databases improve the relevance of search results through several mechanisms that utilize the unique characteristics of vector spaces and machine learning models. Here’s how it works.

Semantic Understanding

As we mentioned before, vector databases store data in high-dimensional vectors. These vectors represent the semantic meaning of the data, capturing nuances that traditional keyword-based searches cannot. A vector database can understand the context and content of a query, enabling it to return results that are not just similar but semantically relevant.

Customizable Similarity Metrics

Vector databases support various similarity metrics. This flexibility allows developers to choose a metric that best matches their definition of relevance during app development. Milvus, for example, supports Euclidean distance, inner product, and cosine similarity for measuring similarity.

Real-World Applications and Case Studies

Tokopedia is one of Indonesia’s largest e-commerce platforms, with an impressive 18 million monthly active users as of 2023, and has become the country’s go-to destination for online shopping. They recognize that their extensive product catalog’s values lie in ensuring customers can discover products based on their preferences quickly.

In the past, Tokopedia utilized Elasticsearch for product search. However, the tools don’t usually consider the semantic meaning of the keywords searched. Instead, it only factors like how often the words appear in the documents and how close they are to each other. One of the solutions that the Tokopedia team found was to find a new way to represent keywords that were later represented as vectors and used Milvus. In the end, one of the reasons Tokopedia chose Milvus is because it offers a broader range of supported indexes. After implementing Milvus as their vector search engine, they achieved an impressive 10x higher click-through rate (CTR) and conversion rate (CVR). You can visit our Tokopedia case study for a more complete story.

Best Practice and Considerations

Best Practices

Data Preparation High-quality data are always important, including for effective search in vector databases. Data preparation begins with data cleaning and preprocessing. Text search may involve tokenization, breaking text into meaningful units, and removing stop words to eliminate common but uninformative words.

Indexing Strategies Index tuning involves several steps: picking the right index type, index parameter tuning, benchmarking, and search parameter tuning. At a higher level, we need to pick indexes based on their type first.

Index Hierarchy.png

Indexes in the hierarchy can handle more queries per second and have better latency, but they can only handle lower data sizes. Instead, indexes in the lower hierarchy can handle more data but can handle fewer queries per second and have longer latency. Other aspects also need to be considered, such as accuracy, index time, and cost. You could refer to this documentation for more information about Milvus-supported indexes and their use cases. There are no obvious choices, as it heavily depends on your use case and the tradeoffs you are willing to take.

Query Optimization Fine-tuning vector models can enhance their ability to capture the nuance of data, improving the relevance of search results. Adjusting query parameters, like selecting the optimal number of nearest neighbors, helps balance precision and performance. To reduce the number of calls, use batch or multi-vector queries instead of multiple single queries.

Challenges and Considerations

Data Privacy Vector databases often contain sensitive information. To ensure compliance with data protection regulations like GDPR and CCPA, anonymize personal data and implement strict access controls. Consider encryption for data at rest and in transit.

Scalability As your application grows, the volume of data and the number of queries can increase significantly. Choose vector databases like Milvus that can be easily distributed. Manage your database or use cloud-based solutions like Zilliz Cloud that dynamically adjust resources based on demand.

There are various approaches that you could take to improve scalability.

Split unrelated data into multiple collections in multiple databases
Database sharding, splitting data based on the hash of the primary key
Partition splitting data based on a filter that is frequently used on that data
Replica, for improving redundancy and more read throughput
Splitting data into segments, marking it as sealed segments (thus read-only), and growing segment

Future of App Search Optimization

Advancements in AI in search are set to elevate vector databases, giving faster and better-aligned searches with user intent. This means that app search will soon offer personalized search results. The future of app search also could become more conversational, just like what already happens with LLMs. Simultaneously, the use of vector databases for searching with images, audio, and video is expected to become more common, revolutionizing how multimedia content is indexed and retrieved.

##Conclusion

Vector databases revolutionize app development by enhancing search functionalities with their ability to perform fast, accurate, and semantic searches. By efficiently handling complex, high-dimensional data, these databases improve the relevance and performance of search results, therefore improving user experience. Embracing vector databases offers a strategic advantage for businesses that aim to enhance their application search capabilities. This approach not only meets the needs of users but also sets a new standard for application development.