Latent Space: The AI Engineer Podcast ⚡️The Rise and Fall of the Vector DB Category
1850 snips
May 1, 2025 
Jo Kristian Bergum, a seasoned search infrastructure expert with two decades at Yahoo and Fast Search & Transfer, dives deep into the evolution of vector databases. He discusses the surge in vector database popularity post-ChatGPT and the misconceptions surrounding embedding-based similarity search. The conversation explores the dynamic interplay between traditional search methods and embedding techniques. Additionally, Joe sheds light on the future of retrieval-augmented generation and the importance of knowledge graphs in AI development.
AI Snips
Chapters
Transcript
Episode notes
Choose Search System by Scale
- Use PostgreSQL with pgvector for moderate scale vector search if you already use it as your database.
- For critical search business needs, consider specialized search engines for better search quality.
Embeddings Blur Search and Recommendations
- Embedding-based retrieval has long been integral to recommender systems and now converges with search technologies.
- There's a layered approach of retrieving candidates and re-ranking them before final presentation.
Build Search with Hybrid Approach
- Start search systems with classical keyword algorithms like BM25 to establish baselines.
- Add embeddings and re-ranking layers as budget and latency allow, tuning sequence by use case.