Your RAG over PDF reports handles prose well and gets n...

When a RAG system fails on numerical data despite performing well on prose, the issue is typically rooted in the ingestion phase. You should inspect the retrieved chunks to see if the table structure was preserved. If the header row and the data rows are separated into different chunks, the retriever will fail to associate the query’s vocabulary with the correct data points. Consequently, the system ranks the labels rather than the actual answer, leading to incorrect retrieval results.