been playing around with how to stop rag from hallucinating on relational data and found some decent ways to handle it. instead of just dumping everything into a vector store, you can try anything from basic row-level chunking to more complex
text-to-sql
routing for structured queries. the key is making sure those tables are actually
retrievable when the user asks something specific. avoid simple character splitting because it totally breaks the context of the cells.
>it's basically a spectrum between semantic search and hard queryingthe hybrid approach is usually the only thing that works for complex schemasanyone else found success with using metadata enrichment to bridge the gap between text and rows?
more here:
https://hackernoon.com/chunking-strategies-for-structured-data-in-rag-systems?source=rss