I came from Data Engineering stuff before jumping into LLM stuff, i am surprised that many people in this space never heard Elastic/OpenSearch

A viral post from a former data engineer is challenging the LLM community's tooling choices, pointing out that many developers have never heard of Elasticsearch or OpenSearch for retrieval-augmented generation (RAG). The core argument is that these established search engines operate on very similar technical principles as specialized vector stores and services like Google Search—the primary difference is one of scale. From an LLM application perspective, both fall under the RAG umbrella, and developers can leverage them for efficient document retrieval without necessarily relying on heavy embedding models.

The post details the technical flexibility these platforms offer. Developers can bypass embeddings entirely and use classic, battle-tested retrieval algorithms like TF-IDF or BM25, which are built into Elastic/OpenSearch. For semantic search, a small BERT model (around 100MB, running on CPU) can be enabled directly within the search engine to generate vector embeddings. This approach is particularly effective for document sets under approximately 10,000 items with good variance. The engineer notes that for deeper semantic tasks or very closely related documents, more powerful embedding models are preferable, but for many practical RAG use cases, Elasticsearch and OpenSearch provide a production-ready, scalable, and often simpler alternative to the newer wave of vector databases.