RAG in 2025: State of the Art and the Road Forward
summary
Argues that RAG beats fine-tuning and naive long-context for enterprise use, cites an average retrieval accuracy of ~80% across ~100 datasets with wide variance (some sets 90%+, some 20–60%), and shows embedding tricks — domain-specific embeddings, hybrid lexical+vector search with re-rankers, quantization and “Matryoshka” subvector truncation — that cut storage by 10x–100x with only single-digit percent drops. Describes deployed features: multimodal embeddings, a re-ranker, and an auto-trunking embedding that returns trunk vectors plus cross-trunk meta for cheaper, more focused LLM context.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker background
- 1:10Why RAG for enterprise use
- 2:35RAG vs long context vs fine-tuning
- 5:28Recent advancements in retrieval accuracy
- 7:54Reducing storage costs with compression
- 8:37Techniques to improve RAG performance
- 10:46Future evolution of RAG systems
- 13:12Simplifying RAG with model layer growth
- 15:21Multimodal embedding for RAG
- 16:44Auto-trunking and context-aware embedding
- 18:07Q&A and closing remarks
related talks
30mRAG: Tricks from the trenches.
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
28mNavigating Challenges and Enhancing Performance of LLM based Applications
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
53mBuild RAG from Scratch
25mAdvanced RAG Systems: Scaling Retrieval-Augmented Generation in Production
45mMaking semantic search & RAG real: How to build a prod-ready app
21mHow RAG is Redefining AI Capabilities