Beyond Text: Building Multimodal Search with pgvector in PostgreSQL
Presented by:
Nasiullha Chaudhari
Nasiullha Chaudhari is a Developer Engagement Manager at YugabyteDB, where he helps engineers build scalable, AI-ready applications on distributed PostgreSQL. He is an international speaker with extensive experience across AI, cloud, DevOps, and Kubernetes, and has delivered talks at conferences around the world.
Nasiullha is a HashiCorp Ambassador and Docker Captain, recognized for his contributions to the developer community. Through his YouTube channel with 180K+ subscribers, he educates developers on AI, cloud-native architectures, and practical workflows for building resilient systems. He also has a 70K+ LinkedIn following and is passionate about helping engineers ship production-ready AI applications, demystifying complex distributed systems, and empowering the developer community to succeed in the AI era.
No video of the event yet, sorry!
Large language models are no longer limited to text. They can also understand images and other modalities and represent them as embeddings that can be stored and queried directly in PostgreSQL. In this talk, we explore how pgvector can be used to build multimodal search features where users search using text, images, or a combination of both, all within a familiar SQL-based workflow.
We begin with the core concepts relevant to PostgreSQL developers: what embeddings are, how multimodal models produce them, and why vector distance functions such as cosine and L2 work well for similarity search. Using a concrete example application—an image- and text-based recipe search—we show how to model embeddings in PostgreSQL using vector columns and how to design schemas that support text-only, image-only, and combined embeddings alongside relational data.
From there, the session walks through practical SQL examples, including inserting embeddings, running nearest-neighbor searches using pgvector operators like <-> and <=>, and combining vector similarity with traditional SQL filters, joins, and ordering. This demonstrates how vector search fits naturally into existing PostgreSQL query patterns rather than replacing them.
The talk concludes with a discussion of performance and indexing considerations as datasets grow. We cover what changes as vector tables scale, how indexing choices affect recall and latency, and what lessons recent work on scalable vector search in PostgreSQL-compatible systems provides for designing production-ready search features. The focus remains on practical trade-offs and rules of thumb that PostgreSQL developers can apply today.
- Date:
- 2026 April 21 16:00 PDT
- Duration:
- 50 min
- Room:
- San Pedro (Level C)
- Conference:
- Postgres Conference: 2026
- Language:
- Track:
- Dev
- Difficulty:
- Medium