@bashifuirkashi: One of the easiest ways to make your RAG system worse is blindly using Top-K retrieval. Top-K just means your retriever grabs the top K highest-scoring chunks and sends them to the LLM. So if K = 5, you always send 5 chunks. The problem? The 5th-best chunk might still be terrible. And now you’re mixing good context with irrelevant or outdated information, which can actually make the final answer worse. A better production approach is to retrieve more candidates, then use things like: ⚙️ Reranking ⚙️ Relevance thresholds ⚙️ Dynamic K ⚙️ Metadata filtering The goal isn’t to always send the top 5. The goal is to send the best context. Comment “RAG” and I’ll send you my AI engineering community with a personalized roadmap, live calls with working AI engineers, recruiters, and hiring managers.

Bashi | Software Engineer

Open In TikTok:

Region: US

Thursday 03 September 2026 15:30:17 GMT