• Background: Vectors of numbers are important for search systems. "Bags of words" approach represents document contents with vectors of word counts. Word embeddings create multidimensional vector spaces for semantic content. Vector databases store and operate on multidimensional vectors.
• Low's Article: Yucheng Low (co-founder of XetHub) argues that a hybrid approach of combining traditional text retrieval algorithms with vector technology is better for "retrieval augmented generation" (RAG) than using a vector database.
• Low's Demonstration: Compares three text retrieval approaches - traditional keyword search with heuristics (BM25), retrieval using just vector embeddings, and first using BM25 to retrieve 1000 results and then picking the best with vector embeddings. Finds the hybrid approach produces better results.
• What's RAG: An approach to improve large language model (LLM) AI systems. Solves problems like providing references and keeping the system on target. Bing's Copilot uses RAG.
• Surprising Conclusion: Today's computers are fast enough to search through thousands of vectors, so it's more efficient to use traditional search technology first and then vector technology to rank results. Also got better RAG results with the hybrid approach. It doesn't mean there's no need for vector databases; just not for RAG.