[Stanley]

RAG is a Data Problem Pretending to Be AI

Retrieval-Augmented Generation fails most often not because LLMs “hallucinate,” but because retrieval pipelines return incomplete, stale, or irrelevant context—due to weak chunking, naive ranking, missing metadata/ACLs, and lack of evaluation—so reliable RAG requires treating it like search + ETL with rigorous instrumentation, hybrid retrieval, rerankers, confidence thresholds, and continuous evals.


https://hackernoon.com/rag-is-a-data-pro...source=rss

[RangoWorld]

Just like back in the 40s and 50s we had revolutionary design and theories for the CPU itself that included the usage of memory chips.

We need similar kind of innovation here with LLM architecture where there's a memory bank that somehow saves and retrieves information at real time like a brain. But storing of information has to be algorithmic and not primarily just space heavy.