Unlocking the future of search: Insights from CTO Jovica Zorić

At our recent Friday Talk, Jovica Zorić, our CTO, took us on an insightful journey Towards smarter search, exploring the evolution of search methodologies—from simple pattern matching to advanced AI-driven techniques.

Jovica shared his hands-on experiences and experiments in developing smarter search functionalities.

Exploring the foundations of search

His exploration began with the Model Context Protocol (MCP), a client-server architecture that enables large language models (LLMs) to interact with various systems. In this case, he focused on using MCP to facilitate seamless interaction with PostgreSQL. This initial exploration triggered a deeper dive into the challenges of creating effective e-commerce search solutions.

To illustrate his points, Jovica used a real-world dataset of UK products from Kaggle, providing a practical foundation for the demonstration. He started with traditional search techniques, beginning with simple wildcard searches using SQL’s “LIKE/ILIKE” operator. While straightforward, this method struggles with variations and typos.

To address these limitations, he introduced fuzzy search using PostgreSQL extensions like pg_trgm, which enhance search accuracy by allowing typo tolerance and approximate matching. However, while fuzzy search improves the handling of misspellings, it doesn’t fully capture the meaning behind a query. This is where full-text search in PostgreSQL comes into play.

Transitioning to AI-powered search

Jovica then explored PostgreSQL’s advanced full-text search capabilities, explaining how they enable more sophisticated term matching and ranking. Considering both the relevance and frequency of words, this approach significantly refines search accuracy.

He also demonstrated how combining fuzzy search with full-text search creates a robust and flexible solution for handling different queries. He also mentioned dedicated search solutions like Elasticsearch, Typesense, and Qdrant. While PostgreSQL provided a solid foundation, these tools offer specialized search capabilities and the best choice ultimately depends on specific requirements.

The talk then transitioned into AI-powered search, where Jovica introduced vector embeddings—a technique that transforms text into numerical representations of semantic meaning. He explained how cosine distance measures similarity between vectors, allowing searches to be based on meaning rather than just keywords.

To implement this, he leveraged the pg_vector extension in PostgreSQL, which enables efficient vector searches. He also showcased an embedding server written in Python, converting search queries into vectors. This allowed for natural language queries like “phone with the biggest battery” to return relevant results based on intent rather than exact wording.

Enhancing search with large language models

Finally, Jovica integrated LLMs to enhance the search process further. He demonstrated how LLMs refine search results by understanding context and intent in natural language queries. For example, when asked about the “best budget camera for wildlife photography,” the LLM could identify and highlight products with specific features, such as high frames per second (FPS) for capturing fast action. This showed how AI-driven search goes beyond traditional methods, delivering greater accuracy and relevance.

Throughout his presentation, Jovica emphasized the importance of choosing the right search technique for the specific use case. From simple pattern matching and dedicated solutions to AI-enhanced search, his demonstrations showcased the strengths and limitations of each approach, highlighting how they work together to create smarter search functionalities.
 

Explore our blog section for upcoming captivating topics that we’ll be sharing with you soon.