LLMs Can't Do It All: Why Embeddings Still Matter

​In the age of powerful large language models, embeddings remain a critical component in retrieval, recommendation, and understanding systems. This session explores why embeddings still matter, how they’re built, and how to choose the right architecture for your domain and application.

​Agenda Highlights:

• ​Context length limitations and the rise of RAG

• ​Why LLMs aren’t ideal for embedding generation

• ​Causal attention vs contrastive learning

• ​Sparse vs dense retrieval (TF-IDF, SPLADE, hybrid approaches)

• ​Bi-Encoders vs Cross-Encoders vs Late Interaction

• ​Latent interaction models and HNSW search

• ​Contrastive learning, triple loss, Siamese networks

• ​Speed vs accuracy trade-offs

• ​Domain-specific embedding tuning (e.g., code vs text)

• ​Evaluation with BEIR and MTEB benchmarks

​Speaker: Sandro Barnabishvili, AI Researcher