Hybrid Search in PostgreSQL: Combining Vector and Full-Text for Real-World Applications
Presented by:
Shayon Sanyal
Shayon is a database architect specializing in PostgreSQL and AI/ML applications. With extensive experience helping enterprises implement vector search, agentic AI systems, and hybrid search architectures, he focuses on practical patterns that work at production scale. A regular speaker at database conferences, Shayon is passionate about making PostgreSQL's advanced capabilities accessible to the broader community.
No video of the event yet, sorry!
PostgreSQL has evolved into a multi-model database capable of handling diverse search requirements. While many developers know about pgvector for semantic search or tsvector for full-text search, few have explored combining these capabilities into unified hybrid search systems. This session takes a practitioner's approach to building hybrid search in PostgreSQL. We'll start with the "why": semantic search excels at meaning but misses exact keywords, while full-text search captures terms but ignores context. Real applications need both. We'll dive into practical implementation patterns:
Understanding the Tradeoffs:
- When vector similarity wins (synonyms, concepts, multilingual)
- When full-text wins (exact matches, Boolean logic, phrases)
- Why combining them outperforms either alone
Schema Design
- Storing embeddings and tsvector columns efficiently
- Choosing embedding dimensions for your use case
- Index strategies for hybrid workloads
Reciprocal Rank Fusion (RRF)
- The algorithm that powers hybrid search at scale
- Implementing RRF in pure SQL using CTEs and window functions
- Tuning the k parameter for your relevance needs
Query Patterns
- Parallel execution of vector and text searches
- Combining scores from different ranking systems
- Filtering before vs. after fusion
Advanced Techniques
- Adding metadata boosting (recency, popularity, user preferences)
- Recursive CTEs for relationship-aware search
- Pre-filtering vectors with text matches for performance
Live demonstrations will show how a single PostgreSQL query can find documents that are semantically similar to a question AND contain specific keywords—returning a unified, ranked result set. We'll examine real query plans, discuss performance characteristics, and cover common pitfalls like score normalization mistakes and inefficient fusion strategies.
Attendees will leave with SQL patterns they can immediately apply, understanding of when hybrid search adds value versus complexity, and appreciation for how PostgreSQL's extensibility makes it uniquely suited for converged search workloads. No proprietary tools required—just PostgreSQL and community extensions.
Key Takeaways:
- Implement Reciprocal Rank Fusion in pure SQL
- Design schemas supporting multi-modal search
- Choose between vector, full-text, or hybrid based on query characteristics
- Avoid common pitfalls in hybrid implementations
- Date:
- Duration:
- 50 min
- Room:
- Conference:
- Postgres Conference: 2026
- Language:
- Track:
- Dev
- Difficulty:
- Medium