RAG Is Simpler Than You Think
Points and comments are a snapshot, not live.
Start with BM25 and query rewriting before jumping to embeddings or vector databases.
The article argues that many teams over-engineer RAG by jumping straight to embeddings. It presents six architectures, from simplest (BM25 full-text search) to most complex (full pre-embedding), and advises moving up the ladder only when data proves it's necessary. Query rewriting with an LLM costs ~$0.001 per query and avoids the complexity of chunking and re-embedding. On-the-fly embedding costs ~$15/month for 1,000 daily queries but adds 200-500ms latency. Hot/cold tiering balances freshness and cost. The author estimates 60% of systems should stop at full-text plus query rewriting.
What commenters are saying
Commenters split on the article's recommendations. Some agree that many teams over-engineer and that keyword search with query rewriting covers most use cases, especially for technical queries. Others argue embeddings are simple to implement and provide significant semantic benefit, making them the right default for document-based RAG. A few criticize the article's style as AI slop with short punchy sentences and buzzwords. Several commenters note that vector embeddings and RAG are just rebrandings of older information retrieval and machine learning techniques.